WO2019083055A1 - 기계학습을 이용한 오디오 복원 방법 및 장치 - Google Patents
기계학습을 이용한 오디오 복원 방법 및 장치Info
- Publication number
- WO2019083055A1 WO2019083055A1 PCT/KR2017/011786 KR2017011786W WO2019083055A1 WO 2019083055 A1 WO2019083055 A1 WO 2019083055A1 KR 2017011786 W KR2017011786 W KR 2017011786W WO 2019083055 A1 WO2019083055 A1 WO 2019083055A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- parameter
- decoding
- unit
- machine learning
- audio signal
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/04—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using predictive techniques
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/04—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using predictive techniques
- G10L19/16—Vocoder architecture
- G10L19/167—Audio streaming, i.e. formatting and decoding of an encoded audio signal representation into a data stream for transmission or storage purposes
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
- G06F3/16—Sound input; Sound output
- G06F3/162—Interface to dedicated audio devices, e.g. audio drivers, interface to CODECs
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/092—Reinforcement learning
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/0017—Lossless audio signal coding; Perfect reconstruction of coded audio signal by transmission of coding error
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
- G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
- G10L19/032—Quantisation or dequantisation of spectral components
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/038—Speech enhancement, e.g. noise reduction or echo cancellation using band spreading techniques
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
Definitions
- the present invention relates to an audio decompression method and apparatus, and more particularly, to an audio decompression method and apparatus for restoring a decoding parameter or an audio signal obtained from a bitstream using machine learning to provide an improved sound quality.
- An audio codec technology capable of transmitting, reproducing and storing high quality contents has been developed. According to the ultra high quality technology, it is possible to transmit, reproduce and store audio of a resolution of 24bit / 192khz.
- the resolution of 24bit / 192khz means that the original audio is sampled at 192kHz and that one sampled signal can be expressed in 2 ⁇ 24 steps using 24 bits.
- AI Artificial intelligence
- AI is a computer system that implements human-level intelligence. Unlike existing Rule-based smart systems, AI is a system in which machines learn, judge and become smart. Artificial intelligence systems are increasingly recognized and improving their understanding of user preferences as they are used, and existing rule-based smart systems are gradually being replaced by deep-run-based artificial intelligence systems.
- Machine learning is an algorithm technology that classifies / learns the characteristics of input data by itself.
- Element technology is a technology that simulates functions such as recognition and judgment of human brain using machine learning algorithms such as deep learning. Understanding, reasoning / prediction, knowledge representation, and motion control.
- Linguistic understanding is a technology for recognizing, applying, and processing human language / characters, including natural language processing, machine translation, dialogue system, query response, speech recognition / synthesis, and the like.
- Visual understanding is a technology for recognizing and processing objects as human vision, including object recognition, object tracking, image search, human recognition, scene understanding, spatial understanding, and image enhancement.
- Inference prediction is a technique for judging and logically inferring and predicting information, including knowledge / probability based reasoning, optimization prediction, preference base planning, and recommendation.
- Knowledge representation is a technology for automating human experience information into knowledge data, including knowledge building (data generation / classification) and knowledge management (data utilization).
- the motion control is a technique for controlling the autonomous travel of the vehicle and the motion of the robot, and includes motion control (navigation, collision, traveling), operation control (behavior control), and the like.
- the present disclosure performs machine learning using various decoding parameters of the original audio and audio codec to obtain reconstructed decoding parameters.
- this scheme can restore higher quality audio using restored decoding parameters.
- the present disclosure provides a method and apparatus for reconstructing a decoding parameter or an audio signal obtained from a bitstream using machine learning.
- An audio decompression method includes decoding a bitstream and obtaining a plurality of decoding parameters for a current frame, generating a plurality of decoding parameters based on a first parameter included in the plurality of decoding parameters, Determining a characteristic of a second parameter included in decryption parameters and associated with a first parameter, applying a machine learning model to at least one of a plurality of decryption parameters, a second parameter, reconstructed second parameter, and decoding the audio signal based on the reconstructed second parameter.
- the step of decoding the audio signal includes the steps of obtaining a corrected second parameter by correcting the restored second parameter based on the characteristic of the second parameter, And decrypting the audio signal based on the second parameter.
- the step of determining the characteristic of the second parameter includes the step of determining a range of the second parameter based on the first parameter, wherein the acquiring step includes acquiring, as the corrected second parameter, a value in a range closest to the restored second parameter when the restored second parameter is not in the range.
- the step of determining a characteristic of a second parameter comprises determining a characteristic of a second parameter based on at least one of a first parameter and a second parameter, And determining a characteristic of the second parameter using the second parameter.
- acquiring the reconstructed second parameter includes: determining candidates of a second parameter based on a property of a second parameter; And selecting one of the two parameter candidates.
- the step of acquiring the reconstructed second parameter acquires the reconstructed second parameter of the current frame based further on at least one of the plurality of reconstructing parameters of the previous frame
- the method comprising the steps of:
- a machine learning model is generated by mechanically learning at least one of an original audio signal and a plurality of decryption parameters.
- an audio decompression method includes decoding a bitstream and obtaining a plurality of decoding parameters for a current frame, decoding an audio signal based on the plurality of decoding parameters, Selecting one of the plurality of machine learning models based on at least one of the decoding parameters and the decoded audio signal and reconstructing the decoded audio signal using the selected machine learning model And a control unit.
- a machine learning model is generated by mechanically learning a decoded audio signal and an original audio signal.
- the step of selecting a machine learning model comprises the steps of determining a start frequency of a bandwidth extension based on at least one of a plurality of decryption parameters, And selecting a machine learning model of the decoded audio signal based on the frequency and the frequency of the decoded audio signal.
- the step of selecting a machine learning model comprises obtaining a gain of a current frame based on at least one of a plurality of decryption parameters, Selecting a machine learning model for a transient signal if the difference between the average of the gain of the current frame and the average of the gains is greater than a threshold value; Determining if the window type included in the plurality of decoding parameters is a short, if the difference value is smaller than a threshold value; selecting a machine learning model for the transient signal when the window type is short, If the type is not short, then the machine learning model for the stationary signal is selected. And a system.
- an audio decompression apparatus includes a memory for storing a received bitstream, and a decoding unit for decoding a bitstream to obtain a plurality of decoding parameters for a current frame and storing the decoding parameters in a plurality of decoding parameters Determining a characteristic of a second parameter included in the plurality of decoding parameters and associated with the first parameter based on the first parameter, and determining at least one of the characteristics of the plurality of decoding parameters, the second parameter, and the second parameter And at least one processor for applying a machine learning model to obtain a reconstructed second parameter and decoding the audio signal based on the reconstructed second parameter.
- At least one processor corrects a restored second parameter based on a characteristic of a second parameter to obtain a corrected second parameter, And decodes the audio signal based on the audio signal.
- At least one processor is configured to use a pre-trained machine learning model based on at least one of a first parameter and a second parameter to generate a second parameter And the characteristic of the light source is determined.
- At least one processor determines the candidates of the second parameter based on the characteristics of the second parameter, and selects one of the candidates of the second parameter based on the machine learning model And obtaining the restored second parameter.
- At least one processor is configured to obtain a reconstructed second parameter of a current frame based further on at least one of a plurality of decoding parameters of a previous frame.
- At least one processor is characterized in that the machine learning model is generated by machine learning at least one of an original audio signal and a plurality of decoding parameters.
- an audio decompression apparatus includes a memory for storing a received bitstream, a decoding unit for decoding a bitstream to obtain a plurality of decoding parameters for a current frame, Decodes the audio signal, selects one of the plurality of machine learning models based on at least one of the plurality of decoding parameters and the decoded audio signal, and decodes the decoded audio signal using the selected machine learning model And at least one processor for reconstructing the image.
- the program for implementing the audio restoration method as described above can be recorded in a computer-readable recording medium.
- FIG. 1 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- FIG. 2 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- FIG. 3 shows a flowchart of an audio decompression method according to an embodiment.
- FIG. 4 shows a block diagram for machine learning in accordance with one embodiment.
- FIG. 5 shows a prediction of the characteristics of the decoding parameters according to an embodiment.
- Figure 6 shows a prediction of the characteristics of a decoding parameter according to an embodiment.
- FIG. 7 shows a flowchart of an audio decompression method according to an embodiment.
- Figure 8 shows the decoding parameters according to an embodiment.
- FIG. 9 illustrates a change in the decoding parameter according to an embodiment.
- FIG. 10 illustrates a change in decoding parameters when the number of bits is increased according to an embodiment.
- FIG. 11 shows a change in the decoding parameter according to an embodiment.
- FIG. 12 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- FIG. 13 shows a flowchart of an audio decompression method according to an embodiment
- FIG. 14 shows a flowchart of an audio restoration method according to an embodiment.
- FIG. 15 shows a flowchart of an audio decompression method according to an embodiment.
- part used in the specification means software or hardware component, and " part " However, “ part " is not meant to be limited to software or hardware. &Quot; Part " may be configured to reside on an addressable storage medium and may be configured to play back one or more processors.
- part (s) refers to components such as software components, object oriented software components, class components and task components, and processes, Subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and variables.
- the functions provided in the components and " parts " may be combined into a smaller number of components and “ parts “ or further separated into additional components and " parts ".
- processor should be broadly interpreted to include a general purpose processor, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a controller, a microcontroller, a state machine, In some circumstances, a “ processor " may refer to an application specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA)
- ASIC application specific integrated circuit
- PLD programmable logic device
- FPGA field programmable gate array
- processor refers to a combination of processing devices, such as, for example, a combination of a DSP and a microprocessor, a combination of a plurality of microprocessors, a combination of one or more microprocessors in conjunction with a DSP core, It can also be called.
- memory should be broadly interpreted to include any electronic component capable of storing electronic information.
- the terminology memory may be any suitable memory such as random access memory (RAM), read-only memory (ROM), non-volatile random access memory (NVRAM), programmable read-only memory (PROM), erase- May refer to various types of processor-readable media such as erasable programmable read-only memory (PROM), flash memory, magnetic or optical data storage devices, registers, and the like.
- RAM random access memory
- ROM read-only memory
- NVRAM non-volatile random access memory
- PROM programmable read-only memory
- erase- May to various types of processor-readable media such as erasable programmable read-only memory (PROM), flash memory, magnetic or optical data storage devices, registers, and the like.
- a memory is said to be in electronic communication with a processor if the processor can read information from and / or write information to the memory.
- the memory integrated in the processor is in electronic communication with the processor.
- FIG. 1 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- the audio decompression apparatus 100 may include a receiving unit 110 and a decoder 120.
- the receiving unit 110 may receive the bit stream.
- the decoding unit 120 may output the decoded audio signal based on the received bitstream.
- the audio restoration apparatus 100 will be described in detail with reference to FIG.
- FIG. 2 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- the audio decompression apparatus 100 may include a codec information derivation unit 210 and at least one decoding unit.
- the codec information derivation unit 210 may correspond to the receiving unit 110 of FIG.
- the at least one decoding unit may include at least one of a first decoding unit 221, a second decoding unit 222, and an Nth decoding unit. At least one of the first decoding unit 221, the second decoding unit 222, and the Nth decoding unit may correspond to the decoding unit 120 of FIG.
- the codec information derivation unit 210 may receive the bitstream.
- the bit stream can be generated in the encoding apparatus.
- the encoding apparatus can compress the original audio into a bit stream.
- the codec information derivation unit 210 may receive the bitstream from the encoding unit or the storage medium through wired / wireless communication.
- the codec information derivation unit 210 may store the bitstream in a memory.
- the codec information derivation unit 210 may extract various information from the bitstream.
- Various information may include codec information.
- the codec information may include information about the technique used for the original audio to be encoded.
- the techniques used to encode the original audio may be MP3, AAC, HE-AAC, and the like.
- the codec information derivation unit 210 may select one of the at least one decoding unit based on the codec information.
- At least one decoding unit may include a first decoding unit 221, a second decoding unit 222, and an Nth decoding unit 223.
- the decoding unit selected by the codec information derivation unit 210 among the at least one decoding unit can decode the audio signal based on the bitstream.
- the N-th decoding unit 223 will be described for convenience of explanation.
- the first decoding unit 221 and the second decoding unit 222 may have a similar structure to the Nth decoding unit 223.
- the N-th decoding unit 223 may include an audio signal decoding unit 230.
- the audio signal decoding unit 230 may include a lossless decoding unit 231, an inverse quantization unit 232, a stereo reconstruction unit 233 and an inverse transformation unit 234.
- the lossless decoding unit 231 can receive the bitstream.
- the lossless decoding unit 231 may decode the bitstream and output at least one decoding parameter.
- the lossless decoding unit 231 can decode the bitstream without loss of information.
- the inverse quantization unit 232 can receive at least one decoding parameter from the lossless decoding unit.
- the inverse quantization unit 232 can dequantize at least one decoding parameter.
- the dequantized decoding parameter may be a mono signal.
- the stereo signal reconstruction unit 233 can reconstruct the stereo signal based on the inverse quantized decoding parameter.
- the inverse transform unit 234 may convert the stereo signal in the frequency domain and output the decoded audio signal in the time domain.
- the decoding parameter may include at least one of a spectral bin, a scale factor gain, a global gain, spectral data, and a window type.
- Decoding parameters may be parameters used in codecs such as MP3, AAC, HE-AAC.
- the decryption parameter is not limited to a specific codec, but may have a decryption parameter having a similar function even if the name is different.
- the decoding parameter may be transmitted on a frame-by-frame basis. A frame is a unit of dividing the original audio signal in the time domain.
- the spectral bin may correspond to the magnitude of the signal along the frequency in the frequency domain.
- the scale factor gain and the global gain are values for scaling the spectral bean.
- the scale factor may have a different value for each of a plurality of bands included in one frame.
- the global gain can have the same value for all bands in one frame.
- the audio decompression apparatus 100 may obtain the audio signal in the frequency domain by multiplying the spectral bin, the scale factor gain, and the global gain.
- the spectral data is information indicating the characteristics of the spectral bean.
- the spectral data may represent the sign of the spectral bean.
- the spectral data may also indicate whether the spectral bin is zero.
- the window type may represent a characteristic of the original audio signal. And may correspond to a time interval for converting the original audio signal of the time domain into the frequency domain. If the original audio signal is a stationary signal with little change, the window type can represent "long”. If the original audio signal is a transient signal with significant changes, the window type may indicate "short".
- the Nth decoding unit 123 may include at least one of the parameter characteristic determination unit 240 and the parameter recovery unit 250.
- the parameter characteristic determination unit 240 may receive at least one decoding parameter and determine the characteristics of the at least one decoding parameter.
- the parameter characteristic determination unit 240 can use the machine learning to determine the characteristics of the at least one decoding parameter.
- the parameter characteristic determination unit 240 may use the first decoding parameter included in the at least one decoding parameter to determine the characteristics of the second decoding parameter included in the at least one decoding parameter.
- the parameter characteristic determination unit 240 may also output at least one of the characteristics of the decoding parameter and the decoding parameter to the parameter restoring unit 250.
- the parameter characteristic determination unit 240 will be described in detail with reference to FIG. 4 to FIG.
- the parameter restoring unit 250 may receive at least one decoding parameter from the lossless decoding unit 231.
- the parameter restoring unit 250 may restore at least one decoding parameter.
- the parameter reconstruction unit 250 may use a mechine learning model to reconstruct at least one decoding parameter.
- the audio signal decoding unit 230 can output a decoded audio signal close to the original audio based on the restored at least one decoding parameter.
- the parameter restoring unit 250 may receive at least one decoding parameter and a characteristic of the decoding parameter from the parameter characteristic determining unit 240.
- the parameter restoring unit 250 may apply the machine learning model to the characteristics of at least one decoding parameter and the decoding parameter to output the restored parameter.
- the parameter restoring unit 250 may apply the machine learning model to at least one decoding parameter to output the restored parameter.
- the parameter restoring unit 250 can correct the restored parameter based on the parameter characteristic.
- the parameter restoring unit 250 may output the corrected parameter.
- the audio signal decoding unit 130 can output a decoded audio signal close to the original audio based on the corrected parameter.
- the parameter restoring unit 250 may output at least one of the restored decoded and corrected parameters to the parameter characteristic determining unit 240 or the parameter restoring unit 250. At least one of the parameter characteristic determination unit 240 and the parameter restoring unit 250 may receive at least one of the decoded parameter and the corrected parameter of the previous frame. The parameter characteristic determination unit 240 may output the parameter characteristic of the current frame based on at least one of the at least one decoding parameter and the corrected parameter of the previous frame. The parameter restoring unit 250 may obtain restored parameters of the current frame based on at least one of the at least one decoded parameter and the corrected parameter of the previous frame.
- FIG. 3 shows a flowchart of an audio decompression method according to an embodiment.
- the audio decompression apparatus 100 may decode the bitstream and obtain a plurality of decoding parameters for the current frame.
- the audio decompression apparatus 100 may determine the characteristics of the second parameter.
- the audio decompression apparatus 100 may obtain the restored second parameter using a machine learning model.
- the audio decompression apparatus 100 may decode the audio signal based on the restored second parameter.
- the audio decompression apparatus 100 may decode the bitstream and obtain a plurality of decoding parameters for the current frame (operation 310).
- the lossless decoding unit 231 can obtain a plurality of decoding parameters by decoding the bitstream.
- the lossless decoding unit 231 can output the decoding parameters to the inverse quantization unit 232, the parameter characteristic determination unit 240, or the parameter restoration unit 250.
- the audio decompression apparatus 100 may analyze the decryption parameter and determine where to output the decryption parameter.
- the audio decompression apparatus 100 may determine where to output the decryption parameters according to a predetermined rule.
- the bitstream may include information on where the decoding parameter is to be output.
- the audio decompression apparatus 100 can determine where to output the decoding parameters based on the information included in the bitstream.
- the audio decompression apparatus 100 may not modify at least one decoding parameter if it can guarantee high sound quality without modifying at least one decoding parameter among the plurality of decoding parameters.
- the lossless decoding unit 231 can output at least one decoding parameter to the inverse quantization unit 232. [ At least one parameter may not be modified because it does not pass through the parameter characteristic determination unit 240 or the parameter restoration unit 250.
- the audio decompression apparatus 100 does not use the parameter characteristic determination unit 240 and the parameter decompression unit 250 for some decoding parameters, and thus can efficiently use computing resources.
- the audio decompression apparatus 100 may determine to modify at least one decoding parameter.
- the lossless decoding unit 231 can output at least one decoding parameter to the parameter restoring unit 250.
- the audio decompression apparatus 100 can obtain the reconstructed decoding parameter based on the decoding parameter using the machine learning model.
- the audio decompression apparatus 100 may decode the audio signal based on the reconstructed decoding parameter.
- the audio restoration apparatus 100 can provide an audio signal of improved sound quality based on the restored decryption parameter.
- the machine learning model will be described in more detail with reference to FIG.
- the audio decompression apparatus 100 may decide to modify a plurality of decryption parameters.
- the lossless decoding unit 231 may output a plurality of decoding parameters to the parameter characteristic determination unit 240.
- the parameter characteristic determination unit 240 may determine a characteristic of the second parameter included in the plurality of decoding parameters based on the first parameter included in the plurality of decoding parameters (operation 320).
- the second parameter may be associated with the first parameter.
- the first parameter may directly or indirectly indicate the characteristics of the second parameter.
- the first parameter may be at least one of scale factor gain, global gain, spectral data and window type for the second parameter.
- the first parameter may be a parameter adjacent to the second parameter. Also, the first parameter may be a parameter included in the same band or frame as the second parameter. The first parameter may be a band including the second parameter or a parameter included in a band or a frame adjacent to the frame.
- the present disclosure distinguishes the first parameter and the second parameter for convenience of explanation, the first parameter may be the same as the second parameter. That is, the parameter characteristic determination unit 240 can determine the characteristics of the second parameter from the second parameter itself.
- the parameter restoring unit 250 may obtain a second parameter reconstructed by applying the machine learning model to at least one of the plurality of decoding parameters, the second parameter, and the second parameter (step 330).
- the audio decompression apparatus 100 may decode the audio signal based on the restored second parameter (step 340).
- the decoded audio signal based on the second parameter restored by applying the machine learning model can provide excellent sound quality.
- the machine learning model will be described in more detail with reference to FIG.
- FIG. 4 shows a block diagram for machine learning in accordance with one embodiment.
- the data learning unit 410 and the data application unit 420 may be performed at different times.
- the data learning unit 410 can operate in advance of the data application unit 420.
- the parameter characteristic determination unit 240 and the parameter recovery unit 250 may include at least one of the data learning unit 410 and the data application unit 420.
- the data learning unit 410 may include a data acquisition unit 411, a preprocessing unit 412, and a machine learning unit 413. [ The data learning unit 410 receives the input data 431 and outputs the machine learning model 432 as a training process.
- the data acquisition unit 411 can receive input data.
- the input data 431 may include at least one of an original audio signal and decoding parameters.
- the original audio signal may be an audio signal recorded in high quality.
- the original audio signal can be expressed in the frequency domain or the time domain.
- the decoding parameters may be the result of encoding the original audio signal. Some information may be lost while encoding the original audio signal. That is, the audio signal decoded from the plurality of decoding parameters may have a lower sound quality than the original audio signal.
- the preprocessing unit 412 can preprocess the acquired data so that the input data 431 can be used for learning.
- the preprocessing section 412 can process input data into a predetermined format so that the machine learning section 413, which will be described later, can use the input data 431.
- the original audio signal and the plurality of decoding parameters have different formats, the original audio signal or a plurality of decoding parameters may be converted to match the format.
- the codec information of the original audio signal and the plurality of decoding parameters may be modified so as to be compatible with each other.
- the original audio signal and a plurality of decoding parameters are represented in different domains, they can be modified to be displayed on the same domain.
- the preprocessing section 412 can select data necessary for learning from the input data 431. [ The selected data may be provided to the machine learning unit 413.
- the preprocessing unit 412 can select data necessary for learning from the preprocessed data according to a preset reference.
- the preprocessing unit 312 can also select data according to a predetermined criterion by learning by the machine learning unit 413, which will be described later.
- the machine learning unit 413 can output the machine learning model 432 based on the selected input data.
- the selected input data may be at least one of a plurality of decoding parameters of the original audio signal.
- the machine learning model 432 may be a criterion for restoring at least one parameter among a plurality of decryption parameters.
- the machine learning unit 413 can learn such that the difference between the audio signal decoded by the reconstructed decoding parameter and the original audio signal is minimized.
- the machine learning unit 413 can learn a criterion as to which input data 431 should be used to restore at least one parameter among a plurality of decryption parameters.
- the machine learning section 413 can learn the machine learning model using the input data 431.
- the machine learning model 432 may be a pre-trained model.
- the machine learning model 432 may be a pre-trained model that receives basic learning data (e.g., at least one decoding parameter).
- the basic learning data may be initial data for building a pre-trained model.
- the machine learning model 432 can be selected in consideration of the application field of the recognition model, the purpose of learning, or the computer performance of the apparatus.
- the machine learning model may be, for example, a model based on a neural network.
- models such as Deep Neural Network (DNN), Recurrent Neural Network (RNN), and Bidirectional Recurrent Deep Neural Network (BRDNN) may be used as a machine learning model.
- DNN Deep Neural Network
- RNN Recurrent Neural Network
- BBDNN Bidirectional Recurrent Deep Neural Network
- the machine learning unit 414 determines a data recognition model that is highly relevant to the input data 431 or the basic learning data, .
- the input data 431 or the basic learning data may be pre-classified according to the type of data, and the data recognition model may be pre-built according to the type of data.
- the input data 431 or the basic learning data may include various types of data such as an area where data is generated, a time at which data was generated, a size of data, a genre of data, It can be classified as a standard.
- the machine learning unit 413 can learn a data recognition model using, for example, a learning algorithm including an error back-propagation method or a gradient descent method.
- the machine learning unit 413 can learn the machine learning model 432 through supervised learning using the input data 431 as an input value, for example. In addition, the machine learning unit 413 learns, for example, the type of data necessary for the situation determination without any further guidance, so that the machine learning unit 413 can learn machine learning The model can be learned. Further, the machine learning unit 413 can learn the machine learning model 432 through reinforcement learning using, for example, feedback as to whether the result of the situation determination based on learning is correct.
- the machine learning unit 413 can perform the machine learning using Equations (1) and (2) as follows.
- x is the selected input data used in the machine learning model
- y is the probability of each candidate
- i is the index of the candidates
- j is the index of the selected input data used in the machine learning model
- I is a weighting matrix
- b is a deflection parameter.
- the machine learning unit 413 can obtain the predicted data using an arbitrary weight W and an arbitrary deflection parameter b.
- the predicted data may be reconstructed decoding parameters.
- the machine learning unit 413 can calculate the cost of y.
- the cost may be the difference between the actual data and the predicted data.
- the cost may be the difference between the data associated with the original audio signal and the data associated with the reconstructed decoding parameter.
- the machine learning unit 413 can update the weight W and the deflection parameter b so that the cost is minimized.
- the machine learning unit 413 can obtain the weight and the deflection parameter at the minimum cost.
- the machine learning unit 413 can represent a weight and a deflection parameter at a minimum cost in a matrix.
- the machine learning unit 413 can acquire the machine learning model 432 using at least one of the weight and the parameter when the cost is minimum.
- the machine learning model 432 may correspond to a matrix of weights and a matrix of parameters.
- the machine learning unit 313 can store the learned machine learning model 432.
- the machine learning unit 413 can store the learned machine learning model 432 in the memory of the data data learning unit 410.
- the machine learning unit 413 may store the learned machine learning model 432 in the memory of the data application unit 420 to be described later.
- the machine learning unit 413 may store the learned machine learning model 432 in an electronic device or in a memory of a server connected via a wired or wireless network.
- the memory in which the learned machine learning model 432 is stored may also store instructions or data associated with, for example, at least one other component of the electronic device.
- the memory may also store software and / or programs.
- the program may include, for example, a kernel, a middleware, an application programming interface (API), and / or an application program (or " application ").
- the model evaluation unit (not shown) inputs the evaluation data to the machine learning model 432, and when the result output from the evaluation data does not satisfy the predetermined criterion, the machine learning unit 413 can re-learn .
- the evaluation data may be predetermined data for evaluating the machine learning model 432.
- the model evaluation unit (not shown) satisfies a predetermined criterion when the number or ratio of evaluation data whose results are not correct out of the results of using the learned machine learning model for evaluation data exceeds a predetermined threshold value It can be evaluated as not successful.
- a predetermined criterion is defined as a ratio of 2%, and the learned machine learning model outputs an incorrect result for evaluation data exceeding 20 out of a total of 1000 evaluation data, 432) may not be appropriate.
- the model evaluating unit can evaluate whether each of the learned machine learning models satisfies a predetermined criterion, and determine a model satisfying a predetermined criterion as a final machine learning model. In this case, when there are a plurality of models satisfying the predetermined criterion, the model evaluation unit can determine any one or a predetermined number of models preset in descending order of the evaluation score as the final machine learning model 432.
- At least one of the data acquisition unit 411, the preprocessing unit 412, the machine learning unit 413 and the model evaluation unit in the data learning unit 410 is fabricated in at least one hardware chip form, .
- at least one of the data acquisition unit 411, the preprocessing unit 412, the machine learning unit 413, and the model evaluation unit may be manufactured in the form of a dedicated hardware chip for artificial intelligence (AI) , Or may be mounted on a variety of electronic devices, such as those manufactured as part of an existing general purpose processor (e.g., a CPU or an application processor) or a graphics-only processor (e.g., a GPU).
- AI artificial intelligence
- the data acquisition unit 411, the preprocessing unit 412, the machine learning unit 413, and the model evaluation unit may be mounted on one electronic device, or may be mounted on separate electronic devices, respectively.
- some of the data acquisition unit 411, the preprocessing unit 412, the machine learning unit 413, and the model evaluation unit may be included in the electronic device, and some of them may be included in the server.
- At least one of the data acquisition unit 411, the preprocessing unit 412, the machine learning unit 413, and the model evaluation unit may be implemented as a software module.
- the at least one software module may be provided by an operating system (OS) or by a predetermined application.
- OS operating system
- OS Operating System
- some of the at least one software module may be provided by an Operating System (OS)
- OS Operating System
- the data data application unit 420 may include a data obtaining unit 421, a preprocessing unit 422, and a result providing unit 423.
- the test process may be that the data data application unit 420 receives the input data 441 and the machine learning model 432 and outputs the output data 442.
- the data acquisition unit 421 can acquire input data.
- the input data 441 may include at least one decoding parameter for decoding the audio signal.
- the preprocessing unit 422 can preprocess the input data 441 so that the input data 441 can be used.
- the preprocessing unit 422 can process the input data 441 into a predetermined format so that the result providing unit 423 to be described later can use the input data 441. [
- the preprocessing unit 422 can select data to be used in the result providing unit 423 from among the preprocessed input data.
- the preprocessing unit 422 can select at least one decoding parameter to be used for improving the sound quality of the audio signal among the preprocessed input data.
- the selected data may be provided to the result provider 423.
- the preprocessing unit 422 may select some or all of the preprocessed input data according to a set criterion for improving the sound quality of the audio signal.
- the preprocessing section 422 can also select data according to a predetermined reference by learning by the machine learning section 413.
- the result provider 423 may apply the data selected by the preprocessor 422 to the machine learning model 432 to output the output data 442.
- the output data 442 may be a reconstructed decoding parameter to provide improved sound quality.
- the audio decompression apparatus 100 can output a decoded audio signal close to the original audio signal based on the restored decoding parameter.
- the result provider 423 may also provide the output data 442 to the preprocessor 422.
- the preprocessor may preprocess the output data 442 and provide it to the result provider 423.
- the output data 442 may be a reconstructed decoding parameter of the previous frame.
- the result deliverer 423 may provide the preprocessor 422 with output data 442 for the previous frame.
- the preprocessing unit 422 may provide the reconstructed decoding parameter of the previous frame to the result providing unit 423 together with the selected decoding parameter of the current frame.
- the result providing unit 423 may generate the output data 442 for the current frame by reflecting the information on the previous frame as well as the restored decoding parameters of the current frame.
- the output data 442 for the current frame may be at least one of a reconstructed decoding parameter or a corrected decoding parameter of the current frame.
- the audio restoration apparatus 100 can provide audio with improved sound quality based on the output data 442 for the current frame.
- the model update unit may cause the machine learning model 432 to be updated based on the evaluation of the output data 342 provided by the result providing unit 423.
- the model updating unit may cause the machine learning unit 413 to update the machine learning model 432 by providing the machine learning unit 413 with the output data 442 provided by the result providing unit 423 .
- At least one of the data obtaining unit 421, the preprocessing unit 422, the result providing unit 423, and the model updating unit in the data application unit 420 may be manufactured in at least one hardware chip form, .
- at least one of the data acquiring unit 421, the preprocessing unit 422, the result providing unit 423, and the model updating unit may be manufactured in the form of a dedicated hardware chip for artificial intelligence (AI) , Or may be mounted on a variety of electronic devices, such as those manufactured as part of an existing general purpose processor (e.g., a CPU or an application processor) or a graphics-only processor (e.g., a GPU).
- AI artificial intelligence
- the data acquiring unit 421, the preprocessing unit 422, the result providing unit 423, and the model updating unit may be mounted on one electronic device or on separate electronic devices, respectively.
- some of the data acquisition unit 421, the preprocessing unit 422, the recognition result providing unit 423, and the model updating unit may be included in the electronic device, and some of the model updating unit may be included in the server.
- At least one of the data obtaining unit 421, the preprocessing unit 422, the result providing unit 423, and the model updating unit may be implemented as a software module.
- the result providing unit 423 and the model updating unit is implemented as a software module (or a program module including an instruction) May be stored in a computer-readable, non-transitory computer readable media.
- the at least one software module may be provided by an operating system (OS) or by a predetermined application.
- OS Operating System
- OS Operating System
- FIG. 1 operations of the audio decompression apparatus 100 of FIG. 1, the data learning unit 410 of FIG. 4, and the data application unit 420 will be described in detail with reference to FIGS. 5 to 11.
- FIG. 1 operations of the audio decompression apparatus 100 of FIG. 1, the data learning unit 410 of FIG. 4, and the data application unit 420 will be described in detail with reference to FIGS. 5 to 11.
- FIG. 1 operations of the audio decompression apparatus 100 of FIG. 1, the data learning unit 410 of FIG. 4, and the data application unit 420 will be described in detail with reference to FIGS. 5 to 11.
- FIG. 1 operations of the audio decompression apparatus 100 of FIG. 1, the data learning unit 410 of FIG. 4, and the data application unit 420 will be described in detail with reference to FIGS. 5 to 11.
- FIG. 5 shows a prediction of the characteristics of the decoding parameters according to an embodiment.
- the parameter characteristic determination unit 240 can determine the characteristics of the decoding parameters.
- the audio decompression apparatus 100 can reduce the amount of computation because it is not necessary to perform processing on parameters that do not satisfy the characteristics of the decryption parameter.
- the audio decompression apparatus 100 can prevent the reconstructed decoding parameter from deriving a sound quality worse than the input decoding parameter.
- the graph 510 may indicate the magnitude of the signal according to the frequency for one frame.
- the plurality of decoding parameters obtained by the audio decompression apparatus 100 based on the bitstream may include a magnitude value of the signal according to the frequency.
- the magnitude value of the signal may correspond to a spectral bin.
- the plurality of decryption parameters may include a first parameter and a second parameter.
- the parameter characteristic determination unit 140 can determine the characteristics of the second parameter based on the first parameter.
- the first parameter may be a parameter adjacent to the second parameter.
- the audio restoration apparatus 100 can determine the characteristics of the second parameter based on the trend of the first parameter.
- the characteristic of the second parameter may be a range of the second parameter.
- the second parameter may be the magnitude value 513 of the signal at frequency f3.
- the first parameter may be the magnitude values 511, 512, 514, and 515 of the signals corresponding to the frequencies f1, f2, f4, and f5.
- the audio decompression apparatus 100 may determine that the magnitude values 511, 12, 514, and 515 of the signals corresponding to the first parameter are upward trends. Therefore, the audio decompression apparatus 100 can determine the range of the magnitude value 513 of the signal corresponding to the second parameter between the signal value 512 and the signal value 514.
- the parameter characteristic determination unit 240 of FIG. 2 may include the data learning unit 410 of FIG.
- the machine learning model 432 may be pre-trained by the data learning unit 410.
- the data learning unit 410 of the parameter characteristic determination unit 240 can receive information corresponding to the original audio signal.
- the information corresponding to the original audio signal may be information obtained by encoding the original audio signal itself or the original audio signal with high quality.
- the data learning unit 410 of the parameter characteristic determination unit 240 can also receive the decoding parameters.
- the parameters received by the data learning unit 410 of the parameter characteristic determination unit 240 may correspond to at least one frame.
- the data learning unit 410 of the parameter characteristic determination unit 240 can also output the machine learning model 432 based on the operations of the data acquisition unit 411, the preprocessing unit 412, and the machine learning unit 413 have.
- the data learning unit 410 of the machine learning model 432 may be a machine learning model 432 for determining the characteristics of the second parameter based on the first parameter.
- the machine learning model 432 may be given as a weight for each of the at least one first parameters.
- the parameter characteristic determination unit 240 may include the data application unit 420 shown in FIG.
- the parameter characteristic determination unit 240 can determine the characteristics of the second parameter based on at least one of the first parameter and the second parameter.
- the parameter characteristic determination unit 240 may use a pre-trained machine learning model to determine the characteristics of the second parameter.
- the data application unit 420 of the parameter property determination unit 240 may receive at least one of the first parameter and the second parameter included in the plurality of decoding parameters of the current frame.
- the data application unit 420 of the parameter property determination unit 240 may receive the machine learning model 432 from the data learning unit 410 of the parameter property determination unit 240.
- the data application unit 420 of the parameter property determination unit 240 can determine the characteristics of the second parameter based on the operations of the data acquisition unit 421, the preprocessing unit 422, and the result provision unit 423.
- the data application unit 420 of the parameter property determination unit 240 may determine the characteristics of the second parameter by applying the machine learning model 432 to at least one of the first parameter and the second parameter.
- the audio decompression apparatus 100 can restore the second parameter not included in the bitstream to provide high bit rate audio.
- the second parameter of the audio decompression apparatus 100 may be the magnitude value of the signal at the frequency f0.
- the bitstream may not include information on the size of the signal at the frequency f0.
- the audio restoration apparatus 100 can estimate the characteristics of the signal at the frequency f0 based on the first parameter.
- the first parameter may be the magnitude values 511, 512, 513, 514, and 515 of the signals corresponding to the frequencies f1, f2, f3, f4, and f5.
- the audio decompression apparatus 100 may determine that the magnitude values 511, 512, 513, 514, and 515 of the signal corresponding to the first parameter are upward trends. Therefore, the audio decompression apparatus can determine the range of the magnitude value of the signal corresponding to the second parameter to be between the signal value 514 and the signal value 515.
- the audio restoration apparatus 100 may include at least one of the data learning unit 410 and the data application unit 420 shown in FIG. Since the operation of the data learning unit 410 or the data application unit 420 has already been described, a detailed description thereof will be omitted here.
- the second parameter may be the magnitude value 523 of the signal at frequency f3.
- the first parameter may be the signal magnitude values 521, 522, 524, and 525 corresponding to the frequencies f1, f2, f4, and f5.
- the audio decompression apparatus 100 can determine that the magnitude values 521, 522, 524, and 525 of the signals corresponding to the first parameter are rising and descending trends. Since the signal value 524 corresponding to the frequency f4 is larger than the signal value 522 corresponding to the frequency f2, the audio decompression apparatus 100 sets the range of the signal size value 523 corresponding to the second parameter to the signal (524). ≪ / RTI >
- the second parameter may be the magnitude value 533 of the signal at frequency f3.
- the first parameter may be signal magnitude values 531, 532, 534, and 535 corresponding to frequencies f1, f2, f4, and f5.
- the audio decompression apparatus 100 can determine that the magnitude values 531, 532, 534, and 535 of the signal corresponding to the first parameter are the rising trend after the falling. Since the signal value 534 corresponding to the frequency f4 is smaller than the signal value 532 corresponding to the frequency f2, the audio decompression apparatus 100 sets the range of the signal size value 533 corresponding to the second parameter to the signal (534). ≪ / RTI >
- the second parameter may be the magnitude value 543 of the signal at frequency f3.
- the first parameter may be the magnitude values 541, 542, 544, and 545 of the signals corresponding to the frequencies f1, f2, f4, and f5.
- the audio decompression apparatus 100 may determine that the magnitude values 541, 542, 544, and 545 of the signals corresponding to the first parameter are falling trends.
- the audio decompression apparatus 100 can determine the range of the magnitude value of the signal corresponding to the second parameter between the signal value 542 and the signal value 544.
- Figure 6 shows a prediction of the characteristics of a decoding parameter according to an embodiment.
- a plurality of frames may be used to determine the characteristics of the decoding parameters for one frame of the audio decompressor 100.
- the audio restoration apparatus 100 may use frames one frame before to determine the characteristics of the decoding parameters for one frame.
- the audio decompression apparatus 100 may be configured to decode at least one of a frame n-2, a frame n-1 610, a frame n 620, or a frame n-1 to determine the characteristics of at least one decoding parameter included in the frame n + At least one decoding parameter included in frame n + 1 630 may be used.
- the audio decompression apparatus 100 can obtain decoding parameters from the bitstream.
- the audio decompression apparatus 100 may obtain the graphs 640, 650, and 660 based on the decoding parameters in a plurality of frames.
- Graph 640 may represent the decoding parameters for frame n-1 610 in the frequency domain.
- the decryption parameters shown in graph 640 may indicate the magnitude of the signal along the frequency.
- the graph 650 may represent the magnitude of the signal for frequency n (620) in the frequency domain.
- the graph 660 may indicate the magnitude of the signal in frequency domain for frequency n + 1 630.
- the audio decompression apparatus 100 may determine the characteristics of the magnitudes of the signals included in the graph 660 based on the magnitudes of the at least one signal included in the graph 640, the graph 650 and the graph 660.
- the audio decompression apparatus 100 is configured to reconstruct a signal 640 included in the graph 660 based on the magnitude of at least one signal included in the graph 640, the graph 650, Gt; 662 < / RTI >
- the audio decompression apparatus 100 can confirm the trend of the signal sizes 641, 642, and 643 of the graph 640.
- the audio decompression apparatus 100 can also confirm trends of the signal sizes 651, 652, and 653 of the graph 650. The trend may be to rise and then descend near f3.
- the audio restoration device 100 may also determine the trend of the graph 660 based on the graph 640 and the graph 650.
- the audio decompression apparatus 100 can also determine that the size 662 of the signal is greater than or equal to the size 661 of the signal and the size 663 of the signal.
- the audio decompression apparatus 100 is configured to reconstruct the audio signal f0 (x, y) included in the graph 660 based on the magnitude of at least one signal included in the graph 640, the graph 650 and the graph 660. [ Can be determined.
- the audio decompression apparatus 100 can confirm trends of signal sizes in the graph 640.
- the audio decompression apparatus 100 can also confirm the trends of the signal sizes of the graph 650. The trend may be to descend near f0.
- the audio restoration device 100 may also determine the trend of the graph 660 based on the graph 640 and the graph 650.
- the audio decompression apparatus 100 can also determine that the magnitude of the signal at f0 is less than or equal to the magnitude of the signal at f4 and greater than or equal to the magnitude of the signal at f5.
- the audio restoration apparatus 100 may include at least one of the data learning unit 410 and the data application unit 420 shown in FIG. Since the operation of the data learning unit 410 or the data application unit 420 has already been described, a detailed description thereof will be omitted here.
- frames prior to one frame may be used to determine the characteristics of the decoding parameters included in one frame of the audio decompression apparatus 100.
- the audio decompression apparatus 100 can determine the characteristics of the signal according to the specific frequency included in the current frame based on the signal according to the specific frequency included in the previous frame.
- the audio decompression apparatus 100 decodes the decoded parameters according to the specific frequency included in the current frame based on the distribution range, average value, intermediate value, median, minimum value, maximum value, Can be determined.
- the audio decompression apparatus 100 may determine the characteristics of the magnitude 662 of the signal contained in the graph 660 based on the magnitude of the at least one signal included in the graph 640 and the graph 650 have.
- the audio decompression apparatus 100 is able to reconstruct the audio signal at the frequency f3 of the graph 660 based on the magnitude 642 of the signal at the frequency f3 of the graph 640 and the magnitude 652 of the signal at the frequency f3 of the graph 650
- the characteristics of the signal size 662 can be determined.
- the characteristics of signal magnitude 662 may be based on distribution range, mean value, median, median, minimum, maximum, deviation, or sign of signal magnitude 642 and signal magnitude 652.
- the audio decompression apparatus 100 can obtain the decoding parameters from the bitstream.
- the decryption parameter may include a second parameter. Further, the characteristic of the second parameter can be determined based on the already determined parameter rather than the decryption parameter.
- the quantization step size may not be included in the decoding parameter.
- the second parameter may correspond to the magnitude of the signal according to the frequency for one frame.
- the magnitude value of the signal may correspond to a spectral bin.
- the audio decompression apparatus 100 can also determine the range of the spectral bin based on the quantization step size.
- the quantization step size is a range of the size of a signal determined by one spectral bin.
- the quantization step size may vary from frequency to frequency. In the audio frequency domain, the quantization step size can be dense. In an area other than the audio frequency domain, the quantization step size may be affected. Thus, knowing the frequency value corresponding to the spectral bin, the quantization step size can be determined.
- the range of spectral bins can also be determined based on the quantization step size.
- the audio decompression apparatus 100 can obtain the decoding parameters from the bitstream.
- the decryption parameter may include a first parameter and a second parameter.
- the characteristic of the second parameter may be determined based on the first parameter.
- the characteristic of the second parameter may be a range of the second parameter.
- the first parameter may include a scale factor and a masking threshold value.
- the quantization step size may be determined based on the scale factor and the masking threshold.
- the scale factor is a value for scaling the spectral bin as described above.
- the scale factor may have a different value for each of a plurality of bands included in one frame.
- the masking threshold is the minimum size of the current signal for the current signal to be heard when noise is present.
- the masking threshold may vary depending on the frequency and type of masker. The masking threshold can also be increased when the frequency of the masker and the current signal is close.
- the current signal may be present at f0, and there may be a masker signal at f1 close to f0.
- the masking threshold at f0 can be determined by the mask of f1. If the magnitude of the current signal at f0 is less than the masking threshold, the current signal may be an inaudible sound. Therefore, the audio decompression apparatus 100 can ignore the current signal at f0 in the encoding or decoding process. On the other hand, if the magnitude of the current signal at f0 is greater than the masking threshold, the current signal may be audible. Therefore, the audio decompression apparatus 100 can not ignore the current signal at f0 in the encoding or decoding process.
- the audio decompression apparatus 100 may set the quantization step size to a smaller value among the scale factor and the masking threshold value.
- the audio decompression apparatus 100 can also determine the range of the spectral bin based on the quantization step size.
- FIG. 7 shows a flowchart of an audio decompression method according to an embodiment.
- the audio decompression apparatus 100 may decode the bitstream to obtain a plurality of decoding parameters of the current frame for decoding the audio signal.
- the audio decompression apparatus 100 may determine the characteristics of the second parameter included in the plurality of decryption parameters based on the first parameter included in the plurality of decryption parameters.
- the audio decompression apparatus 100 may use the machine learning model to obtain the restored second parameter based on at least one of the plurality of decryption parameters.
- the audio decompression apparatus 100 may correct the second parameter based on the characteristics of the second parameter to obtain the corrected second parameter.
- the audio decompression apparatus 100 may decode the audio signal based on the corrected second parameter.
- Steps 710 and 750 may be performed by the audio signal decoding unit 230.
- Step 720 may be performed by the parameter characteristic determination unit 240.
- steps 730 to 740 may be performed by the parameter restoring unit 250.
- the data learning unit 410 and the data application unit 420 of the parameter restoring unit 250 may receive the characteristics of the second parameter as an input. That is, the parameter restoring unit 250 can perform the machine learning on the basis of the characteristic of the second parameter.
- the data learning unit 410 of the parameter restoring unit 250 may output the machine learning model 432 by reflecting the characteristic of the second parameter.
- the data application unit 420 of the parameter restoring unit 250 may output the output data 442 by reflecting the characteristics of the second parameter.
- the data learning unit 410 and the data application unit 420 of the parameter restoring unit 250 may not receive the property of the second parameter as an input. That is, the parameter restoring unit 250 only performs the machine learning based on the decryption parameter, and may not perform the machine learning on the basis of the characteristics of the second parameter.
- the data learning unit 410 of the parameter restoring unit 250 can output the machine learning model 432 without reflecting the characteristics of the second parameter.
- the data application unit 420 of the parameter restoring unit 250 may output the output data 442 without reflecting the characteristics of the second parameter.
- the output data 442 may be a restored second parameter.
- the parameter restoring unit 250 can determine whether the restored second parameter is suitable for the characteristic of the second parameter. If the restored second parameter matches the characteristic of the second parameter, the parameter restoring unit 250 may output the restored parameter to the audio signal decoding unit 230. [ If the restored second parameter does not match the characteristic of the second parameter, the parameter restoring unit 250 may correct the restored second parameter based on the characteristic of the second parameter to obtain the corrected second parameter. The parameter restoring unit 250 may output the corrected parameter to the audio signal decoding unit 230. [
- the characteristic of the second parameter may be a range of the second parameter.
- the audio restoration apparatus 100 can determine the range of the second parameter based on the first parameter.
- the audio decompression apparatus 100 may obtain a value of a range closest to the restored second parameter as a corrected second parameter when the restored second parameter is not within the range of the second parameter. This will be described in more detail with reference to FIG.
- Figure 8 shows the decoding parameters according to an embodiment.
- the graph 800 shows the magnitude of the signal according to the frequency of the original audio signal in the frequency domain.
- the graph 800 may correspond to one frame of the original audio signal.
- the original audio signal appears as a curve 805 with a continuous waveform.
- the original audio signal can be sampled at frequencies f1, f2, f3 and f4.
- the magnitude of the original audio signal can be represented by dots 801, 802, 803, and 804.
- the original audio signal can be encoded.
- the audio decompression apparatus 100 can generate a decoding parameter by decoding the encoded original audio signal.
- Graph 810 shows the magnitude of the signal along the frequency.
- the dotted line 815 shown in the graph 810 may correspond to the original audio signal.
- the points 811, 812, 813, and 814 shown in the graph 810 may correspond to decryption parameters.
- the decoding parameter may be output from the lossless decoding unit 231 of the audio decompression apparatus 100. [ At least one of the original audio signal and decoding parameters may be scaled and displayed in graph 810.
- the dotted line 815 may be different from the points 811, 812, 813, 814.
- the difference between the dotted line 815 and the dots 811, 812, 813 and 814 may be due to the error caused by encoding and decoding of the original audio signal.
- the audio decompression apparatus 100 can determine the characteristics of the decoding parameters corresponding to the dots 811, 812, 813 and 814. [ The audio decompression apparatus 100 may use a machine learning model to determine the characteristics of the decryption parameters. The determination of the decryption parameter characteristics has already been described with reference to FIG. 5 and FIG. 6, and a detailed description thereof will be omitted.
- the decoding parameter may be a spectral bin.
- the characteristic of the decoding parameter may also be a range of spectral bins.
- a range of spectral bins determined by the audio restoration apparatus 100 may be represented as a graph 830.
- the arrow mark 835 indicates a possible range of the point 831 corresponding to the spectral bin.
- the arrow mark 836 represents the possible range of the point 832 corresponding to the spectral bin.
- the arrow mark 837 indicates the possible range of the point 833 corresponding to the spectral bin.
- the arrow mark 838 represents the possible range of the point 834 corresponding to the spectral bin.
- the audio restoration apparatus 100 can determine the characteristics of the signal at f0 between f2 and f3. The audio restoration apparatus 100 may not receive the decryption parameter for f0. The audio restoration apparatus 100 can determine the characteristics of the decryption parameters at f0 based on the decryption parameters associated with f0.
- the audio decompression apparatus 100 may not receive information related to the size of the spectral bean at f0.
- the audio decompression apparatus 100 can determine the range of the magnitude of the signal at f0 using the spectral bin of the frequency adjacent to f0 and the spectral bin of the frame adjacent to the current frame. This is described in detail with reference to FIG. 5 and FIG. 6, and a detailed description thereof will be omitted.
- the audio restoration apparatus 100 may restore the decoding parameters.
- the audio restoration apparatus 100 can use a machine learning model.
- the audio decompression apparatus 100 may apply at least one of a decoding parameter and a decoding parameter to the machine learning model.
- the restored decoding parameters of the audio decompression apparatus 100 are the same as those of the graph 850.
- the dots 851, 852, 853, 854, and 855 represent restored decoding parameters.
- the reconstructed decoding parameter may have a larger error than the decoding parameter before reconstruction. For example, point 834 corresponding to a spectral bin in graph 830 is close to the original audio signal, but point 854 corresponding to a spectral bin in graph 850 is far from the original audio signal 860 Can be.
- the audio restoration apparatus 100 can correct the decoding parameter.
- the audio decompression apparatus 100 can determine whether the decryption parameter is within a possible range of the decryption parameter.
- the audio decompression apparatus 100 can correct the decryption parameter when the decryption parameter is not within the possible range of the decryption parameter.
- the corrected decoding parameter may be within a possible range of the decoding parameter.
- graph 870 represents the corrected spectral bean.
- the points 871, 872, 873, 875 corresponding to the spectral bin may be within the possible range of the spectral bean. However, the point 874 corresponding to the spectral bin may be outside the possible range 878 of the spectral bin.
- the audio restoration apparatus 100 obtains the value of the nearest spectral bin 878 as the corrected spectral bin .
- the audio decompression apparatus 100 resets the maximum value of the range 878 to a point corresponding to the corrected spectral bin when the point 874 corresponding to the restored spectral bin is a value larger than the maximum value of the range 878 880). That is, the audio decompression apparatus 100 may correct the point 874 corresponding to the restored spectral bin to a point 880.
- Point 880 may correspond to the corrected spectral bin.
- the audio decompression apparatus 100 can decode the audio signal based on the corrected decoding parameter.
- the sampling rate of the audio signal can be improved by the point 875 corresponding to the spectral bin recovered at the frequency f0.
- the point 880 corresponding to the spectral bin reconstructed at the frequency f4 the size of the audio signal can be accurately represented. Since the corrected decoding parameter is close to the original audio signal in the frequency domain, the decoded audio signal may be close to the original audio signal.
- FIG. 9 illustrates a change in the decoding parameter according to an embodiment.
- Graph 910 illustrates graph 810 of FIG.
- Graph 910 shows the magnitude of the signal along the frequency.
- the dotted line 915 shown in the graph 910 may correspond to the original audio signal.
- the points 911, 912, 913, and 914 shown in the graph 910 may correspond to decryption parameters. At least one of the original audio signal and the decoding parameters may be scaled and displayed in the graph 910.
- the audio decompression apparatus 100 can determine the characteristics of the decryption parameters corresponding to the dots 911, 912, 913, and 914.
- the audio decompression apparatus 100 may use a machine learning model to determine the characteristics of the decryption parameters. The determination of the characteristics of the decoding parameters has already been described with reference to FIG. 5 and FIG. 6, and a detailed description thereof will be omitted.
- the decoding parameter may be a spectral bin.
- the characteristic of the decoding parameter may also be a range of spectral bins. The range of spectral bins determined by the audio decompression apparatus 100 may appear as a graph 930. [
- the audio decompression apparatus 100 can determine the candidates for finely adjusting the spectral bin.
- the audio decompression apparatus 100 may represent the spectral bin using a plurality of bits. Also, as the number of bits for expressing the spectral bin increases, the audio decompression apparatus 100 can finely represent the spectral bin.
- the audio decompression apparatus 100 may increase the number of bits for expressing the spectral bean in order to finely adjust the spectral bin. The case of increasing the number of bits for expressing the spectral bin will be described with reference to FIG.
- FIG. 10 illustrates a change in decoding parameters when the number of bits is increased according to an embodiment.
- the audio decompression apparatus 100 may use two bits to represent the quantized decoding parameters.
- the audio decompression apparatus 100 can display the quantization decoded parameters using '00', '01', '10' and '11'. That is, the size of the decoding parameter that can be represented by the audio decompression apparatus 100 is four.
- the audio decompression apparatus 100 may assign the minimum value that the decoding parameter may have to '00'. Also, the audio decompression apparatus 100 can allocate the maximum value that the decoding parameter can have to '11'.
- the size of the decoding parameter received by the audio decompression apparatus 100 may be the same as the point 1020.
- the size of the decryption parameter may be '01'.
- the actual size of the decoded parameters before quantization may be the same as the asterisks 1011, 1012, and 1013.
- the error range may be the same as the arrow 1031.
- the error range may be the same as the arrow 1032.
- the error range may be the same as the arrow 1033.
- the audio decompression apparatus 100 may use three bits to represent the quantized decryption parameters.
- the audio decompression apparatus 100 can quantize the decoded parameters using '000', '001', '010', '011', '100', '101', '110', and '111' have. That is, the size of the decoding parameter that the audio restoration apparatus 100 can display is eight.
- the audio restoration apparatus 100 may assign the minimum value that the decoding parameter may have to '000'. Also, the audio decompression apparatus 100 may allocate the maximum value that the decoding parameter can have to '111'.
- the size of the decoding parameter received by the audio decompression apparatus 100 may be the same as the points 1071, 1072, and 1073.
- the sizes of the decoding parameters may be '001', '101', and '011', respectively.
- the actual size of the decryption parameter may be the same as the asterisks 1061, 1062, 1063. If the actual size of the decryption parameter is the same as the asterisk 1061, the error range may be the same as the arrow 1081. [ If the actual size of the decryption parameter is the same as the asterisk 1062, the error range may be the same as the arrow 1082. [ If the actual size of the decoding parameter is the same as the asterisk 1063, the error range may be the same as the arrow 1083.
- the error of the decoding parameter displayed on the graph 1050 is relatively smaller than the error of the decoding parameter displayed on the graph 1000.
- the decoding parameters can be expressed finely.
- the audio decompression apparatus 100 may determine candidates for finely adjusting the decoding parameters.
- the audio decompression apparatus may additionally use one bit to represent the decoding parameter.
- candidates 951, 952, and 953 corresponding to one decoding parameter 931 of the graph 930 can be determined.
- the audio decompression apparatus 100 may use the characteristics of the decryption parameter to determine the decryption parameter candidates 951, 952, 953.
- the property of the decoding parameter may be a range 954 of the decoding parameter.
- the candidates 951, 952, 953 may be within the range 954 of the decryption parameter.
- the audio restoration apparatus 100 can select one of the decryption parameter candidates 951, 952, and 953 based on the machine learning model.
- the audio restoration apparatus 100 may include at least one of a data learning unit 410 and a data application unit 420.
- the audio decompression apparatus 100 may apply at least one of the decoding parameter of the current frame and the decoding parameter of the previous frame to the machine learning model to select one of the decoding parameters.
- the machine learning model can be pre-trained.
- the decryption parameter may include a first parameter and a second parameter.
- the audio decompression apparatus 100 may use the first parameter associated with the second parameter to select one of the candidates of the second parameter.
- the audio decompression apparatus 100 may obtain the selected decoding parameter 961.
- the audio decompression apparatus 100 can also obtain the decoded audio signal based on the selected decryption parameter 961.
- the audio decompression apparatus may additionally use 2 bits to represent decoding parameters. It is also possible to determine the candidates 971, 972, 973, 974, 975 corresponding to one decoding parameter 931 of the graph 930. The candidates 971, 972, 973, 974, and 975 have finer values than the candidates 951, 952, and 953 of the graph 950.
- the audio decompression apparatus 100 can recover accurate decoding parameters when using 2 bits rather than 1 bit.
- the audio decompression apparatus 100 may use the characteristics of the decryption parameters to determine the decryption parameter candidates 971, 972, 973, 974, 975.
- the property of the decoding parameter may be the range 977 of the decoding parameter.
- Candidates 971, 972, 973, 974, 975 may be within a range 976 of decryption parameters.
- the audio decompression apparatus 100 can select one of the decryption parameter candidates 971, 972, 973, 974, and 975 based on the machine learning model.
- the audio decompression apparatus 100 may apply at least one of the decoding parameter of the current frame and the decoding parameter of the previous frame to the machine learning model to select one of the decoding parameters.
- the decryption parameter may include a first parameter and a second parameter.
- the audio decompression apparatus 100 may use the first parameter associated with the second parameter to select one of the candidates of the second parameter.
- the audio decompression apparatus 100 may obtain the selected decryption parameter 981.
- the selected decryption parameter 981 may be a more accurate value than the selected decryption parameter 961 of the graph 960. That is, the selected decoding parameter 981 may be closer to the dotted line corresponding to the original audio signal than the selected decoding parameter 961.
- the audio decompression apparatus 100 may also obtain the decoded audio signal based on the selected decoding parameter 981.
- FIG. 11 shows a change in the decoding parameter according to an embodiment.
- the audio restoration apparatus 100 may receive the bit stream.
- the audio decompression apparatus 100 can obtain decoding parameters based on the bitstream.
- the audio restoration apparatus 100 can determine the characteristics of the decryption parameter.
- the characteristic of the decoding parameter may be a sign.
- the decoding parameter may have a size of 0, and the size of 0 may be a characteristic of a decoding parameter.
- the decoding parameter may be spectral data.
- the spectral data may represent the sign of the spectral bin.
- the spectral data may also indicate whether the spectral bin is zero.
- Spectral data may be included in the bitstream.
- the audio restoration apparatus 100 can also generate spectral data based on the bitstream.
- the decryption parameter may include a first parameter and a second parameter.
- the audio restoration apparatus 100 can determine the characteristics of the second parameter based on the first parameter.
- the first parameter may be spectral data.
- the second parameter may be a spectral bin.
- the graph 1110 indicates the size of the decoding parameter according to the frequency.
- the decoding parameter may be a spectral bin.
- the decoding parameters may have various codes.
- the decoding parameter 1111 may have a negative sign.
- the decryption parameter 1113 may have a positive sign.
- the audio restoration apparatus 100 may determine the sign of the decryption parameter as the characteristics of the decryption parameters 1111 and 1113.
- Decryption parameter 1112 may have a size of zero.
- the audio decompression apparatus 100 can determine the size of 0 as a characteristic of the decryption parameter 1112. [
- the audio decompression apparatus 100 may apply decryption parameters to the machine learning model to determine decrypted decryption parameters.
- the graph 1130 indicates the size of the reconstructed decoding parameter according to the frequency.
- the audio decompression apparatus 100 may restore the decoding parameters 1111, 1112, and 1113 to obtain the decompressed decoding parameters 1131, 1132, and 1133.
- the reconstructed decoding parameters 1131 and 1133 may have different signs from the decoding parameters 1111 and 1113.
- the recovered decoding parameter 1132 may have a value other than 0, unlike the decoding parameter 1112.
- the audio decompression apparatus 100 can obtain the corrected decoding parameter by correcting the reconstructed decoding parameter based on the characteristics of the decoding parameter.
- the audio decompression apparatus 100 can correct the reconstructed decoding parameter based on the sign of the decoding parameter. Referring to the graph 1150, the audio decompression apparatus 100 can obtain the corrected decoding parameters 1151 and 1153 by correcting the signs of the reconstructed decoding parameters 1131 and 1133.
- the audio decompression apparatus 100 can also obtain the corrected decoding parameter 1152 by correcting the size of the reconstructed decoding parameter 1132 to zero.
- the audio decompression apparatus 100 can obtain the reconstructed decoding parameters by applying a machine learning model to the characteristics of the decoding parameters and the decoding parameters. That is, the audio decompression apparatus 100 may obtain the reconstructed parameters according to the graph 1150 based on the decoding parameters according to the graph 1110.
- FIG. 12 shows a block diagram of an audio decompression apparatus 100 according to an embodiment.
- the audio decompression apparatus 100 may include a codec information derivation unit 1210, an audio signal decoding unit 1220, a bitstream analyzing unit 1230, a decompression method selection unit 1240, and at least one decompression unit.
- the codec information derivation unit 1210 may correspond to the reception unit 110 of FIG.
- the codec information derivation unit 1210 may correspond to the codec information derivation unit 210 of FIG.
- the codec information derivation unit 1210 may receive the bitstream and determine which technique the bitstream is encoded using.
- the techniques used to encode the original audio may be MP3, AAC, HE-AAC, and the like.
- the audio signal decoding unit 1220 decodes the audio signal based on the bit stream.
- the audio signal decoding unit 1220 may correspond to the audio signal decoding unit 230 of FIG.
- the audio signal decoding unit 1220 may include a lossless decoding unit, an inverse quantization unit, a stereo signal restoration unit, and an inverse transform unit.
- the audio signal decoding unit 1220 can output the restored audio signal based on the codec information received from the codec information derivation unit 1210.
- the bitstream analyzing unit 1230 can obtain the decoding parameters for the current frame based on the bitstream.
- the bitstream analyzing unit 1230 can recognize the characteristics of the restored audio signal based on the decoding parameters.
- the bitstream analyzing unit 1230 can transmit information on the characteristics of the signal to the restoration method selection unit 1240.
- the decoding parameter may include at least one of a spectral bin, a scale factor gain, a global gain, a window type, a buffer level, Temporal Noise Shaping (TNS) information, and Perceptual Noise Substitution (PNS) information.
- a spectral bin a scale factor gain
- a global gain a global gain
- a window type a window type
- a buffer level a temporary noise Shaping
- PPS Perceptual Noise Substitution
- the spectral bin may correspond to the magnitude of the signal along the frequency in the frequency domain.
- the audio encoding apparatus can transmit accurate spectral beans only for a human-sensitive frequency range in order to reduce data. In addition, it is possible to transmit a spectral bean or an inaccurate spectral bean for a high-frequency region which is hard to be heard by human beings or a low-frequency region which is hard for human beings to hear.
- the audio decompression apparatus 100 may apply a bandwidth extension technique to an area to which the spectral bin is not transmitted.
- the bitstream analyzing unit 1230 can analyze the spectral bin and determine the frequency region in which the spectral bin is correctly transmitted and the frequency region in which the spectral bin is incorrectly transmitted.
- the bitstream analyzing unit 1230 can transmit information on the frequency to the restoration method selecting unit 1240.
- bandwidth extension techniques can generally be applied to high frequency regions.
- the bitstream analyzer 1230 can determine the minimum frequency value of the frequency domain in which the spectral bin was transmitted incorrectly as the start frequency.
- the bitstream analyzer 1230 may determine that the bandwidth extension technique should be applied from the start frequency.
- the bitstream analyzing unit 1230 can transmit the start frequency to the restoration method selecting unit 1240.
- the scale factor gain and the global gain are values for scaling the spectral bean.
- the bitstream analyzing unit 1230 can analyze the scale factor gain and the global gain to obtain the characteristics of the reconstructed audio signal. For example, the bitstream analyzing unit 1230 can determine that the current frame is a transient signal when the scale factor gain and the global gain of the current frame change abruptly. Also, the bitstream analyzer 1230 can determine that the frames are stationary signals when there is little change in the scale factor gain and the global gain of the frames. The bitstream analyzing unit 1230 can transmit information on whether the frames are a stationary signal or a transient signal to the restoration method selecting unit 1240. [
- the bitstream analyzing unit 1230 can determine that the current frame is a stationary signal when the window type of the current frame indicates " long. &Quot; The bitstream analyzing unit 1230 can determine that the current frame is a transient signal when the window type of the current frame indicates " short. &Quot; The bitstream analyzing unit 1230 can transmit information on whether the frames are a stationary signal or a transient signal to the restoration method selecting unit 1240.
- the buffer level is information on the size of the available bits remaining after coding the frame.
- the buffer level is used when coding using Variable Bit Rate (VBR). If the frame of the original audio is a stationary signal with no change, the audio encoding apparatus can encode the original audio using fewer bits. However, if the original audio frame is a complicated transient signal, the audio encoding apparatus can encode the original audio using many bits.
- the audio coding apparatus has residual bits that are obtained by coding the stationary signal, and can be used for coding the transient signal at a later time. That is, a high buffer level of the current frame means that the current frame is a stationary signal. Also, a low buffer level of the current frame means that the current frame is a transient signal.
- the bitstream analyzing unit 1230 can transmit information on whether the frames are a stationary signal or a transient signal to the restoration method selecting unit 1240.
- Temporal Noise Shaping (TNS) information is information for reducing pre-echo.
- TNS Temporal Noise Shaping
- the start position of an attack signal in the time domain can be known.
- the attack signal means a loud sound that suddenly appears. Since the bitstream analyzing unit 1230 can know the start position of the attack signal by TNS, the bitstream analyzing unit 1230 can determine that it is a stationary signal before the start position. Further, the bitstream analyzing unit 1230 can determine that it is a transient signal after the start position
- the Perceptual Noise Substitution (PNS) information indicates information on a hole in the frequency domain.
- a hole refers to a portion where a spectral bin is not transmitted to save bits of a bit stream, and is filled with random noise upon decoding.
- the bit stream analyzing unit 1230 may transmit information on the position of the hole to the restoration method selecting unit 1240.
- the restoration method selection unit 1240 may receive the decoded audio signal and the characteristics of the decoded parameter.
- the restoration method selection unit 1240 can select a method for restoring the decoded audio signal.
- the audio signal decoded by one of the at least one decompression unit may be reconstructed based on the selection of the reconstruction method selection unit 1240.
- the at least one decompression unit may include a first decompression unit 1250, a second decompression unit 1260, and an Nth decompression unit. At least one of the first reconstruction unit 1250, the second reconstruction unit 1260, and the Nth reconstruction unit may use a machine learning model.
- the machine learning model may be a model generated by mechanically learning at least one of an original audio signal, a decoded audio signal, and a decoding parameter.
- At least one of the first decompression unit 1250, the second decompression unit 1260 and the Nth decompression unit may include a data acquisition unit 1251, a preprocessor 1252, and a result provision unit 1253.
- At least one of the first restoration unit 1250, the second restoration unit 1260, and the Nth restoration unit may include the data learning unit 410 of FIG.
- At least one of the first decompression unit 1250, the second decompression unit 1260 and the Nth decompression unit may receive at least one of the decoded audio signal and the decoding parameter.
- the characteristics of the decoded parameter may be information on the frequency region in which the spectral bin is correctly transmitted and the frequency region in which the spectral bin is transmitted incorrectly.
- the reconstruction method selection unit 1240 can determine to restore the decoded audio signal based on at least one of the decoded parameter and the decoded audio signal.
- the restoration method selection unit 1240 may determine to restore the decoded audio signal using the first restoration unit 1250.
- the first reconstruction unit 1250 can output the reconstructed audio signal using a machine learning model.
- the reconstruction method selection unit 1240 may determine to restore the audio signal using the bandwidth extension technique.
- Bandwidth extension techniques include Spectral Band Replication (SBR).
- the restoration method selection unit 1240 may determine to restore the decoded audio signal using the second restoration unit 1260. [ The second reconstruction unit 1260 can output the reconstructed audio signal using the band extension technique improved by the machine learning model.
- the characteristic of the decoded parameter may be information on whether the frame is a stationary signal or a transient signal. If the frame is a stationary signal, the restoration method selection unit 1240 may use the first restoration unit 1250 for the stationary signal. If the frame is a transient signal, the reconstruction method selection unit 1240 may use the second reconstruction unit 1260 for the transient signal. The first reconstructing unit 1250 or the second reconstructing unit 1260 may output the reconstructed audio signal.
- the characteristic of the decoded parameter may be information on the position of the hole.
- the decompression method selection unit 1240 can determine to restore the decoded audio signal based on the decoded parameter and the decoded audio signal.
- the restoration method selection unit 1240 may determine to restore the decoded audio signal using the first restoration unit 1250.
- the first reconstruction unit 1250 can output the reconstructed audio signal using a machine learning model.
- the reconstruction method selection unit 1240 may use the second reconstruction unit 1260 for the signals of the positions of the holes.
- the second reconstruction unit 1260 can output the reconstructed audio signal using a machine learning model.
- the restoration method selecting unit 1240 can select a method of restoring the decoded audio signal according to the characteristics of the audio signal, so that the audio restoring apparatus 100 can efficiently recover the audio signal.
- FIG. 13 shows a flowchart of an audio decompression method according to an embodiment
- the audio decompression apparatus 100 decodes the bitstream to obtain a plurality of decoding parameters for the current frame.
- the audio decompression apparatus 100 decodes the audio signal based on the plurality of decoding parameters.
- the audio decompression apparatus 100 selects one of the plurality of machine learning models based on at least one of the plurality of decoding parameters and the decoded audio signal.
- the audio decompression apparatus 100 reconstructs the decoded audio signal using the selected machine learning model.
- the audio restoration apparatus 100 according to FIG. 13 and the audio restoration apparatus 100 according to FIG. 3 have a common point in that they can improve the soundness of the decoded audio signal.
- the audio decompression apparatus 100 shown in FIG. 13 is less dependent on the decoding parameter, the versatility can be high.
- FIG. 14 shows a flowchart of an audio restoration method according to an embodiment.
- the codec information derivation unit 1210 can receive the bitstream.
- the audio signal decoding unit 1220 can output the decoded audio signal based on the bit stream.
- the bitstream analyzing unit 1230 can acquire the characteristics of the decoding parameters based on the bitstream. For example, the bitstream analyzer 1230 may determine a start frequency of a bandwidth extension based on at least one of a plurality of decoding parameters (step 1410).
- the audio encoding apparatus can accurately transmit a spectral bin for a frequency region smaller than the frequency f. However, since the frequency region larger than the frequency (f) is a region that is difficult for human beings to hear, the audio encoding apparatus can transmit the spectral bean poorly or not.
- the codec information derivation unit 1210 can determine the start frequency f of the bandwidth extension based on the spectral bin. The codec information derivation unit 1210 may output the start frequency f of the bandwidth extension to the reconstruction method selection unit 1240.
- the restoration method selection unit 1240 can select a machine learning model of the decoded audio signal based on the start frequency and the frequency of the decoded audio signal.
- the restoration method selection unit 1240 may compare the frequency of the decoded audio signal with the start frequency f (step 1420). In addition, the restoration method selecting unit 1240 can select the decoding method based on the comparison.
- the restoration method selection unit 1240 can select a predetermined machine learning model.
- the predetermined machine learning model can be pre-trained by the decoded audio signal and the original audio signal.
- the audio restoration apparatus 100 may restore the decoded audio signal using the machine learning model (step 1430).
- the restoration method selection unit 1240 may determine to restore the decoded audio signal using the bandwidth extension technique. For example, the restoration method selection unit 1240 may select a machine learning model to which the bandwidth extension technique is applied. The machine learning model may be pre-trained using at least one of a parameter associated with the bandwidth extension technique, a decoded audio signal, and an original audio signal. The audio restoration apparatus 100 may restore the decoded audio signal using the machine learning model to which the bandwidth extension technique is applied (operation 1440).
- FIG. 15 shows a flowchart of an audio decompression method according to an embodiment.
- the codec information derivation unit 1210 can receive the bitstream.
- the audio signal decoding unit 1220 can output the decoded audio signal based on the bit stream.
- the bitstream analyzing unit 1230 can acquire the characteristics of the decoding parameters based on the bitstream. For example, the bitstream analyzing unit 1230 may obtain the gain A of the current frame based on at least one of the plurality of decoding parameters (Step 1510). In addition, the bitstream analyzing unit 1230 may obtain an average of the gains of the current frame and frames adjacent to the current frame (step 1520).
- the restoration method selection unit 1240 may compare the difference value between the current frame gain and the average value of the gains with a threshold value (step 1530). Also, the restoration method selection unit 1240 may select a machine learning model for a transient signal when the difference between the gain of the current frame and the average value of the gains is greater than the threshold value. Also, the audio restoration apparatus 100 may restore the decoded audio signal using the machine learning model for the transient signal (step 1550).
- the restoration method selection unit 1240 can determine whether the window type included in the plurality of decoding parameters is short when the difference between the gain of the current frame and the average value of the gains is smaller than the threshold value ). If the window type is short, the restoration method selection unit 1240 may select a machine learning model for the transient signal (step 1550). The restoration method selection unit 1240 can select a machine learning model for a stationary signal when the window type is not short. The audio decompression apparatus 100 may restore the decoded audio signal using a machine learning model for the stasis signal (step 1560).
- the machine learning model for the transient signal can be machine-learned based on the original audio signal and the decoded audio signal classified as transients.
- the machine learning model for the stationary signal can be machine learned based on the original audio signal and the decoded audio signal classified as stationary. Since the stationary signal and the transient signal are different from each other in characteristics, the audio restoration apparatus 100 separately learns the stationary signal and the transient signal, and thus can more efficiently decode the decoded audio signal.
- the above-described embodiments of the present invention can be embodied in a general-purpose digital computer that can be embodied as a program that can be executed by a computer and operates the program using a computer-readable recording medium.
- the computer-readable recording medium includes a storage medium such as a magnetic storage medium (e.g., ROM, floppy disk, hard disk, etc.), optical reading medium (e.g., CD ROM,
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Acoustics & Sound (AREA)
- Signal Processing (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Quality & Reliability (AREA)
- Medical Informatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
Description
Claims (19)
- 비트스트림을 복호화하여(decoding) 현재 프레임에 대한 복수의 복호화 파라미터들을 획득하는 단계;상기 복수의 복호화 파라미터들에 포함되는 제 1 파라미터에 기초하여, 상기 복수의 복호화 파라미터들에 포함되며 상기 제 1 파라미터와 연관되는 제 2 파라미터의 특성을 결정하는 단계;상기 복수의 복호화 파라미터들, 상기 제 2 파라미터 및 상기 제 2 파라미터의 특성 중 적어도 하나에 기계학습모델을 적용하여 복원된(reconstructed) 제 2 파라미터를 획득하는 단계; 및상기 복원된 제 2 파라미터에 기초하여 오디오 신호를 복호화하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 1 항에 있어서,상기 오디오 신호를 복호화하는 단계는,상기 제 2 파라미터의 특성에 기초하여 상기 복원된 제 2 파라미터를 보정하여 보정된 제 2 파라미터를 획득하는 단계; 및상기 보정된 제 2 파라미터에 기초하여 오디오 신호를 복호화하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 2 항에 있어서,상기 제 2 파라미터의 특성을 결정하는 단계는,상기 제 1 파라미터에 기초하여, 상기 제 2 파라미터의 범위를 결정하는 단계를 포함하고,상기 보정된 제 2 파라미터를 획득하는 단계는,상기 복원된 제 2 파라미터가 상기 범위에 있지 않을 경우, 상기 복원된 제 2 파라미터와 가장 가까운 상기 범위의 값을 보정된 제 2 파라미터로 획득하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 1 항에 있어서,상기 제 2 파라미터의 특성을 결정하는 단계는,상기 제 1 파라미터 및 상기 제 2 파라미터 중 적어도 하나에 기초하여 프리-트레이닝된(pre-trained) 기계학습모델을 이용하여 상기 제 2 파라미터의 특성을 결정하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 1 항에 있어서,상기 복원된 제 2 파라미터를 획득하는 단계는,상기 제 2 파라미터의 특성에 기초한 제 2 파라미터의 후보들을 결정하는 단계; 및상기 기계학습모델에 기초하여 상기 제 2 파라미터의 후보들 중 하나를 선택하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 1 항에 있어서,상기 복원된 제 2 파라미터를 획득하는 단계는,이전 프레임의 복수의 복호화 파라미터들 중 적어도 하나에 더 기초하여 상기 현재 프레임의 상기 복원된 제 2 파라미터를 획득하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 1 항에 있어서,상기 기계학습모델은 오리지널 오디오 신호 및 상기 복수의 복호화 파라미터들 중 적어도 하나를 기계학습하여 생성되는 것을 특징으로 하는 오디오 복원 방법.
- 비트스트림을 복호화하여(decoding) 현재 프레임에 대한 복수의 복호화 파라미터들을 획득하는 단계;상기 복수의 복호화 파라미터들에 기초하여 오디오 신호를 복호화하는 단계;상기 복수의 복호화 파라미터들 중 적어도 하나 및 상기 복호화된 오디오 신호에 기초하여 복수의 기계학습모델들 중 하나의 기계학습모델을 선택하는 단계; 및상기 선택된 기계학습모델을 이용하여 상기 복호화된 오디오 신호를 복원하는(reconstructing) 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 8 항에 있어서,상기 기계학습모델은 상기 복호화된 오디오 신호 및 오리지널 오디오 신호를 기계학습하여 생성되는 것을 특징으로 하는 오디오 복원 방법.
- 제 8 항에 있어서,상기 기계학습모델을 선택하는 단계는,상기 복수의 복호화 파라미터들 중 적어도 하나에 기초하여 대역폭 확장(band width extension)의 시작 주파수를 결정하는 단계; 및상기 시작 주파수 및 상기 복호화된 오디오 신호의 주파수에 기초하여 상기 복호화된 오디오 신호의 기계학습모델을 선택하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 제 8 항에 있어서,상기 기계학습모델을 선택하는 단계는,상기 복수의 복호화 파라미터들 중 적어도 하나에 기초하여 현재 프레임의 게인을 획득하는 단계;상기 현재 프레임 및 상기 현재 프레임에 인접한 프레임들의 게인들의 평균을 획득하는 단계;상기 현재 프레임의 게인과 상기 게인들의 평균값의 차이값이 임계값보다 큰 경우 트랜지언트(transient) 신호를 위한 기계학습모델을 선택하는 단계;상기 현재 프레임의 게인과 상기 게인들의 평균값의 차이값이 임계값보다 작은 경우, 상기 복수의 복호화 파라미터들에 포함된 윈도우 타입이 쇼트(short)인지를 결정하는 단계;상기 윈도우 타입이 쇼트인 경우 상기 트랜지언트 신호를 위한 기계학습모델을 선택하는 단계; 및상기 윈도우 타입이 쇼트가 아닌 경우, 스테이셔너리(stationary) 신호를 위한 기계학습모델을 선택하는 단계를 포함하는 것을 특징으로 하는 오디오 복원 방법.
- 수신된 비트스트림을 저장하는 메모리; 및상기 비트스트림을 복호화하여(decoding) 현재 프레임에 대한 복수의 복호화 파라미터들을 획득하고, 상기 복수의 복호화 파라미터들에 포함되는 제 1 파라미터에 기초하여, 상기 복수의 복호화 파라미터들에 포함되며 상기 제 1 파라미터와 연관되는 제 2 파라미터의 특성을 결정하고, 상기 복수의 복호화 파라미터들, 상기 제 2 파라미터 및 상기 제 2 파라미터의 특성 중 적어도 하나에 기계학습모델을 적용하여 복원된(reconstructed) 제 2 파라미터를 획득하고, 상기 복원된 제 2 파라미터에 기초하여 오디오 신호를 복호화하는 적어도 하나의 프로세서를 포함하는 것을 특징으로 하는 오디오 복원 장치.
- 제 12 항에 있어서,상기 적어도 하나의 프로세서는,상기 제 2 파라미터의 특성에 기초하여 상기 복원된 제 2 파라미터를 보정하여 보정된 제 2 파라미터를 획득하고, 상기 보정된 제 2 파라미터에 기초하여 오디오 신호를 복호화하는 것을 특징으로 하는 오디오 복원 장치.
- 제 12 항에 있어서,상기 적어도 하나의 프로세서는,상기 제 1 파라미터 및 상기 제 2 파라미터 중 적어도 하나에 기초하여 프리-트레이닝된(pre-trained) 기계학습모델을 이용하여 상기 제 2 파라미터의 특성을 결정하는 것을 특징으로 하는 오디오 복원 장치.
- 제 12 항에 있어서,상기 적어도 하나의 프로세서는,상기 제 2 파라미터의 특성에 기초한 제 2 파라미터의 후보들을 결정하고, 상기 기계학습모델에 기초하여 상기 제 2 파라미터의 후보들 중 하나를 선택하여 상기 복원된 제 2 파라미터를 획득하 는 단계를 포함하는 것을 특징으로 하는 오디오 복원 장치.
- 제 12 항에 있어서,상기 적어도 하나의 프로세서는,이전 프레임의 복수의 복호화 파라미터들 중 적어도 하나에 더 기초하여 상기 현재 프레임의 상기 복원된 제 2 파라미터를 획득하는 것을 특징으로 하는 오디오 복원 장치.
- 제 12 항에 있어서,상기 적어도 하나의 프로세서는,상기 기계학습모델은 오리지널 오디오 신호 및 상기 복수의 복호화 파라미터들 중 적어도 하나를 기계학습하여 생성되는 것을 특징으로 하는 오디오 복원 장치.
- 수신된 비트스트림을 저장하는 메모리; 및상기 비트스트림을 복호화하여(decoding) 현재 프레임에 대한 복수의 복호화 파라미터들을 획득하고, 상기 복수의 복호화 파라미터들에 기초하여 오디오 신호를 복호화하고, 상기 복수의 복호화 파라미터들 중 적어도 하나 및 상기 복호화된 오디오 신호에 기초하여 복수의 기계학습모델들 중 하나의 기계학습모델을 선택하고, 상기 선택된 기계학습모델을 이용하여 상기 복호화된 오디오 신호를 복원하는(reconstructing) 적어도 하나의 프로세서를 포함하는 것을 특징으로 하는 오디오 복원 장치.
- 제 1 항 또는 제 8 항 중 어느 한 항의 방법을 실행하기 위한 컴퓨터 프로그램을 기록한 컴퓨터로 판독 가능한 기록매체.
Priority Applications (5)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP17929628.0A EP3667663B1 (en) | 2017-10-24 | 2017-10-24 | Audio reconstruction method and device which use machine learning |
| KR1020207006359A KR102551359B1 (ko) | 2017-10-24 | 2017-10-24 | 기계학습을 이용한 오디오 복원 방법 및 장치 |
| US16/652,759 US11545162B2 (en) | 2017-10-24 | 2017-10-24 | Audio reconstruction method and device which use machine learning |
| PCT/KR2017/011786 WO2019083055A1 (ko) | 2017-10-24 | 2017-10-24 | 기계학습을 이용한 오디오 복원 방법 및 장치 |
| CN201780095363.XA CN111164682B (zh) | 2017-10-24 | 2017-10-24 | 使用机器学习的音频重建方法和设备 |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/KR2017/011786 WO2019083055A1 (ko) | 2017-10-24 | 2017-10-24 | 기계학습을 이용한 오디오 복원 방법 및 장치 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2019083055A1 true WO2019083055A1 (ko) | 2019-05-02 |
Family
ID=66247840
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/KR2017/011786 Ceased WO2019083055A1 (ko) | 2017-10-24 | 2017-10-24 | 기계학습을 이용한 오디오 복원 방법 및 장치 |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US11545162B2 (ko) |
| EP (1) | EP3667663B1 (ko) |
| KR (1) | KR102551359B1 (ko) |
| CN (1) | CN111164682B (ko) |
| WO (1) | WO2019083055A1 (ko) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021032719A1 (en) * | 2019-08-20 | 2021-02-25 | Dolby International Ab | Multi-lag format for audio coding |
| WO2021172053A1 (ja) * | 2020-02-25 | 2021-09-02 | ソニーグループ株式会社 | 信号処理装置および方法、並びにプログラム |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11416742B2 (en) * | 2017-11-24 | 2022-08-16 | Electronics And Telecommunications Research Institute | Audio signal encoding method and apparatus and audio signal decoding method and apparatus using psychoacoustic-based weighted error function |
| US11106729B2 (en) * | 2018-01-08 | 2021-08-31 | Comcast Cable Communications, Llc | Media search filtering mechanism for search engine |
| US11501787B2 (en) * | 2019-08-22 | 2022-11-15 | Google Llc | Self-supervised audio representation learning for mobile devices |
| US11609947B2 (en) | 2019-10-21 | 2023-03-21 | Comcast Cable Communications, Llc | Guidance query for cache system |
| US20210199503A1 (en) * | 2019-12-26 | 2021-07-01 | Industrial Technology Research Institute | Data processing system disposed on sensor and method thereof |
| CN111653285B (zh) * | 2020-06-01 | 2023-06-30 | 北京猿力未来科技有限公司 | 丢包补偿方法及装置 |
| KR102334390B1 (ko) * | 2020-06-24 | 2021-12-01 | 연세대학교 산학협력단 | 인공지능 기술을 활용한 코덱 압축 효율 향상 장치 및 방법 |
| CN111798866B (zh) * | 2020-07-13 | 2024-07-19 | 商汤集团有限公司 | 音频处理网络的训练及立体声重构方法和装置 |
| WO2022081915A1 (en) * | 2020-10-15 | 2022-04-21 | Dolby Laboratories Licensing Corporation | Method and apparatus for processing of audio using a neural network |
| CN115691521B (zh) * | 2021-07-29 | 2026-03-13 | 华为技术有限公司 | 一种音频信号的编解码方法和装置 |
| CN113838477B (zh) * | 2021-09-13 | 2024-08-02 | 上海兆言网络科技有限公司 | 音频数据包的丢包恢复方法、装置、电子设备及存储介质 |
| KR102618581B1 (ko) | 2023-03-30 | 2023-12-27 | 엘아이지넥스원 주식회사 | 초고해상도 레이더를 위한 노치 대역의 신호를 추정하기 위한 장치 및 그 방법 |
| US12380902B2 (en) | 2023-10-18 | 2025-08-05 | Cisco Technology, Inc. | Vector quantizer correction for audio codec system |
| US12308037B2 (en) | 2023-10-18 | 2025-05-20 | Cisco Technology, Inc. | Reduced multidimensional indices compression for audio codec system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20040047361A (ko) * | 2002-11-29 | 2004-06-05 | 삼성전자주식회사 | 적은 계산량으로 고주파수 성분을 복원하는 오디오 디코딩방법 및 장치 |
| US20150269953A1 (en) * | 2012-10-16 | 2015-09-24 | Audiologicall, Ltd. | Audio signal manipulation for speech enhancement before sound reproduction |
| WO2016040885A1 (en) * | 2014-09-12 | 2016-03-17 | Audience, Inc. | Systems and methods for restoration of speech components |
| US9640194B1 (en) * | 2012-10-04 | 2017-05-02 | Knowles Electronics, Llc | Noise suppression for speech processing based on machine-learning mask estimation |
Family Cites Families (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100446594B1 (ko) | 1997-04-15 | 2005-06-02 | 삼성전자주식회사 | 음성선스펙트럼주파수의부호화/복호화장치및그방법 |
| US7620546B2 (en) * | 2004-03-23 | 2009-11-17 | Qnx Software Systems (Wavemakers), Inc. | Isolating speech signals utilizing neural networks |
| US7885819B2 (en) * | 2007-06-29 | 2011-02-08 | Microsoft Corporation | Bitstream syntax for multi-process audio decoding |
| KR101513028B1 (ko) | 2007-07-02 | 2015-04-17 | 엘지전자 주식회사 | 방송 수신기 및 방송신호 처리방법 |
| WO2009038402A1 (en) | 2007-09-21 | 2009-03-26 | Lg Electronics Inc. | Digital broadcasting system and data processing method in the digital broadcasting system |
| JP2011518345A (ja) * | 2008-03-14 | 2011-06-23 | ドルビー・ラボラトリーズ・ライセンシング・コーポレーション | スピーチライク信号及びノンスピーチライク信号のマルチモードコーディング |
| KR101461774B1 (ko) | 2010-05-25 | 2014-12-02 | 노키아 코포레이션 | 대역폭 확장기 |
| US9311923B2 (en) | 2011-05-19 | 2016-04-12 | Dolby Laboratories Licensing Corporation | Adaptive audio processing based on forensic detection of media processing history |
| US8793557B2 (en) * | 2011-05-19 | 2014-07-29 | Cambrige Silicon Radio Limited | Method and apparatus for real-time multidimensional adaptation of an audio coding system |
| US9135920B2 (en) | 2012-11-26 | 2015-09-15 | Harman International Industries, Incorporated | System for perceived enhancement and restoration of compressed audio signals |
| CN105229731B (zh) * | 2013-05-24 | 2017-03-15 | 杜比国际公司 | 根据下混的音频场景的重构 |
| KR102251833B1 (ko) * | 2013-12-16 | 2021-05-13 | 삼성전자주식회사 | 오디오 신호의 부호화, 복호화 방법 및 장치 |
| CN104751849B (zh) | 2013-12-31 | 2017-04-19 | 华为技术有限公司 | 语音频码流的解码方法及装置 |
| US20160093308A1 (en) * | 2014-09-26 | 2016-03-31 | Qualcomm Incorporated | Predictive vector quantization techniques in a higher order ambisonics (hoa) framework |
| CN104318927A (zh) * | 2014-11-04 | 2015-01-28 | 东莞市北斗时空通信科技有限公司 | 一种抗噪声的低速率语音编码方法及解码方法 |
| CN105070293B (zh) * | 2015-08-31 | 2018-08-21 | 武汉大学 | 基于深度神经网络的音频带宽扩展编码解码方法及装置 |
-
2017
- 2017-10-24 KR KR1020207006359A patent/KR102551359B1/ko active Active
- 2017-10-24 US US16/652,759 patent/US11545162B2/en active Active
- 2017-10-24 EP EP17929628.0A patent/EP3667663B1/en active Active
- 2017-10-24 CN CN201780095363.XA patent/CN111164682B/zh active Active
- 2017-10-24 WO PCT/KR2017/011786 patent/WO2019083055A1/ko not_active Ceased
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20040047361A (ko) * | 2002-11-29 | 2004-06-05 | 삼성전자주식회사 | 적은 계산량으로 고주파수 성분을 복원하는 오디오 디코딩방법 및 장치 |
| US9640194B1 (en) * | 2012-10-04 | 2017-05-02 | Knowles Electronics, Llc | Noise suppression for speech processing based on machine-learning mask estimation |
| US20150269953A1 (en) * | 2012-10-16 | 2015-09-24 | Audiologicall, Ltd. | Audio signal manipulation for speech enhancement before sound reproduction |
| WO2016040885A1 (en) * | 2014-09-12 | 2016-03-17 | Audience, Inc. | Systems and methods for restoration of speech components |
Non-Patent Citations (2)
| Title |
|---|
| See also references of EP3667663A4 * |
| VOLODYMYR KULESHOV: "AUDIO SUPER-RESOLUTION USING NEURAL NETS", THE 5TH INTERNATIONAL CONFERENCE ON LEARNING REPRESENTATIONS (ICLR 2017, 2 August 2017 (2017-08-02), XP080951067 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021032719A1 (en) * | 2019-08-20 | 2021-02-25 | Dolby International Ab | Multi-lag format for audio coding |
| EP4485457A1 (en) * | 2019-08-20 | 2025-01-01 | Dolby International AB | Multi-lag format for audio coding |
| US12223968B2 (en) | 2019-08-20 | 2025-02-11 | Dolby International Ab | Multi-lag format for audio coding |
| WO2021172053A1 (ja) * | 2020-02-25 | 2021-09-02 | ソニーグループ株式会社 | 信号処理装置および方法、並びにプログラム |
| US12149911B2 (en) | 2020-02-25 | 2024-11-19 | Sony Group Corporation | Signal processing apparatus, signal processing method, and program |
Also Published As
| Publication number | Publication date |
|---|---|
| US20200234720A1 (en) | 2020-07-23 |
| KR20200062183A (ko) | 2020-06-03 |
| CN111164682B (zh) | 2025-07-04 |
| EP3667663A1 (en) | 2020-06-17 |
| KR102551359B1 (ko) | 2023-07-04 |
| US11545162B2 (en) | 2023-01-03 |
| EP3667663C0 (en) | 2024-07-17 |
| EP3667663B1 (en) | 2024-07-17 |
| CN111164682A (zh) | 2020-05-15 |
| EP3667663A4 (en) | 2020-09-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2019083055A1 (ko) | 기계학습을 이용한 오디오 복원 방법 및 장치 | |
| WO2020231230A1 (en) | Method and apparatus for performing speech recognition with wake on voice | |
| WO2020111676A1 (ko) | 음성 인식 장치 및 방법 | |
| WO2020101263A1 (en) | Electronic apparatus and method for controlling thereof | |
| WO2021225256A1 (ko) | 전자 장치 및 이의 제어 방법 | |
| WO2018124620A1 (en) | Method and device for transmitting and receiving audio data | |
| EP3818518A1 (en) | Electronic apparatus and method for controlling thereof | |
| WO2020235712A1 (ko) | 콘텐츠 기반의 스타일을 갖는 텍스트 또는 음성을 생성하는 인공 지능 장치 및 그 방법 | |
| WO2019194451A1 (ko) | 인공지능을 이용한 음성 대화 분석 방법 및 장치 | |
| WO2012036487A2 (en) | Apparatus and method for encoding and decoding signal for high frequency bandwidth extension | |
| WO2013141638A1 (ko) | 대역폭 확장을 위한 고주파수 부호화/복호화 방법 및 장치 | |
| WO2020230926A1 (ko) | 인공 지능을 이용하여, 합성 음성의 품질을 평가하는 음성 합성 장치 및 그의 동작 방법 | |
| WO2021029642A1 (en) | System and method for recognizing user's speech | |
| WO2019124963A1 (ko) | 음성 인식 장치 및 방법 | |
| WO2020145472A1 (ko) | 화자 적응형 모델을 구현하고 합성 음성 신호를 생성하는 뉴럴 보코더 및 뉴럴 보코더의 훈련 방법 | |
| WO2016024853A1 (ko) | 음질 향상 방법 및 장치, 음성 복호화방법 및 장치와 이를 채용한 멀티미디어 기기 | |
| WO2020130262A1 (ko) | 컴퓨팅 장치 및 그 동작 방법 | |
| WO2018174397A1 (ko) | 전자 장치 및 제어 방법 | |
| WO2014185569A1 (ko) | 오디오 신호의 부호화, 복호화 방법 및 장치 | |
| WO2020130260A1 (en) | Mobile terminal and method of operating the same | |
| WO2019231068A1 (en) | Electronic device and control method thereof | |
| WO2022092569A1 (ko) | 전자 장치 및 이의 제어 방법 | |
| WO2015093742A1 (en) | Method and apparatus for encoding/decoding an audio signal | |
| EP3545519A1 (en) | Method and device for transmitting and receiving audio data | |
| WO2021040490A1 (en) | Speech synthesis method and apparatus |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 17929628 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 20207006359 Country of ref document: KR Kind code of ref document: A |
|
| ENP | Entry into the national phase |
Ref document number: 2017929628 Country of ref document: EP Effective date: 20200310 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| WWG | Wipo information: grant in national office |
Ref document number: 201780095363.X Country of ref document: CN |