JPH06295196A - Speech recognition device and signal recognition device - Google Patents

Abstract

(57) [Summary] (Correction) [Purpose] A voice recognition device and a signal recognition device capable of accurately detecting a voice signal with a small amount of calculation. [Structure] The speech recognition apparatus 10 learns from a feature extraction unit 14 for detecting a voice signal, a noise signal and a voice signal that are classified into several types in advance, and features parameters sent from the feature extraction unit 14. A voice signal detection unit 15 for discriminating noise and a voice signal by a neural network calculation based on the above, and a feature parameter for all frames sent from the feature extraction unit 14 is subjected to a predetermined calculation to perform a feature parameter for recognition. Feature extraction unit 16 for calculating
And a recognition unit 17 that performs a calculation based on the characteristic parameter sent from the characteristic extraction unit 16 to recognize a voice and output a recognition result, and a display unit 1 that displays the recognition result from the recognition unit 17.
8 and a control unit 19 for controlling the operation of each unit, and a voice signal detection unit 15 using a neural network detects a voice section by a neural network of noise signals and voice signals.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a signal recognition device for recognizing a voice or the like, and more particularly to a voice recognition device having a voice signal detection function and capable of accurately determining a voice signal. The present invention relates to a signal recognition device.

[0002]

2. Description of the Related Art A voice recognition device is classified into a specific speaker system that requires registration of a person's voice before use and an unspecified speaker system that anyone can use without registration. In the specific speaker voice recognition, the factors of individual differences that make voice recognition difficult can be ignored, so that a high-performance voice recognition device can be easily realized.

As a conventional voice recognition device of this type, a discrete word recognition device for recognizing a voice uttered by dividing it into word units is known. In this discrete word recognition device, the voice input from the microphone is converted into a time series of spectrum parameters by the voice analysis unit. First, in the registration mode, the analysis result of the voice uttered as the standard pattern is cut out in units of words by voice detection, and is stored in the standard pattern memory for each word together with the category code. Next, in the recognition mode, the analysis result of the input voice is similarly cut out for each word, and the matching unit performs matching processing with each standard pattern (template) to calculate the matching amount (distance). The determination unit displays / outputs the category code of the standard pattern with the smallest distance as the recognition result.

As the voice analysis, a spectrum analysis by band-pass filters of 12 to 16 channels is often used. The analysis result is sampled at a frame rate of 10 to 20 ms, normalized by amplitude for each frame, and expressed with an accuracy of about 8 bits. Other analysis methods include cepstrum analysis, mel cepstrum analysis, LPC cepstrum analysis, and LPC analysis. For example, when a voice is analyzed and parameters are extracted, in order to eliminate the influence of the utterance length, as shown in FIG. 6, the number of frames is determined and the overlap of the frames is adjusted for analysis.

[0005]

However, in such a conventional voice recognition device, since the detection of the voice section is performed by the voice level or the power level, the noise level is high or the voice level is high. If the level is low, the voice section cannot be detected, or the detected position is inaccurate. With the same analysis as recognition, there are problems that the amount of calculation is too large and the hardware becomes large, or that it cannot be processed in real time.

Therefore, an object of the present invention is to provide a voice recognition device and a signal recognition device which can accurately detect a voice signal with a small amount of calculation.

[0007]

The invention according to claim 1 is
To achieve the above object, a voice data storage means for storing voice data, a voice signal detection means for detecting a noise signal and a voice signal of the voice data stored in the voice data storage means, and a detection by the voice signal detection means Frame dividing means for dividing the voice section of the audio signal thus obtained into frames, characteristic parameter extracting means for extracting characteristic parameters from the frame data divided by the frame dividing means, and characteristic parameters extracted by the characteristic parameter extracting means And a recognition means for recognizing voice data.

According to a second aspect of the present invention, there is provided a voice data storage means for storing voice data, and a first feature for extracting a characteristic parameter used by the voice signal detection means from the voice data stored in the voice data storage means. Parameter extracting means, audio signal detecting means for detecting a noise signal and an audio signal based on the characteristic parameter output from the first characteristic parameter extracting means, and an audio section of the audio signal detected by the audio signal detecting means. And a second characteristic parameter extracting means for extracting a characteristic parameter for voice recognition from the frame data divided by the frame dividing means,
And recognition means for recognizing voice data using the characteristic parameters extracted by the characteristic parameter extracting means.

The voice signal detecting means may be, for example, the device of claim 3.
Alternatively, the audio signal detecting means may be configured by a neural network as described in claim 4, and may be of a predetermined type. Learning may be performed by the classified noise signal and voice signal, and the noise signal and the voice signal may be discriminated by neural network calculation.

According to a fifth aspect of the present invention, data storage means for storing signals inputted in time series as data, signal detection means for detecting noise and signals of the data stored in the data storage means, and Frame dividing means for dividing the delimiter section of the signal detected by the signal detecting means into frames, characteristic parameter extracting means for extracting characteristic parameters from the frame data divided by the frame dividing means, and extraction by the characteristic parameter extracting means And a recognition means for recognizing the time-series data by using the generated characteristic parameter.

The audio signal detecting means may be, for example, the device of claim 6.
, It may be configured by a neural network.

[0012]

According to the invention described in claims 1, 2, 3 and 4, the voice is digitized and captured by the voice capturing means, and the captured voice data is stored in the voice data storage means. The voice data stored in the voice data storage means is the first
The characteristic parameter extracting means extracts the characteristic parameter for voice signal detection, and the noise signal and the voice signal are detected by the voice signal detecting means using the neural network based on the characteristic parameter. When the noise signal and the voice signal are detected, the voice section of the voice signal detected by the frame dividing means is divided into frames. In this case, the voice section is determined by analyzing the voice data and determining the beginning and end of the word.

The voice division of the voice signal is divided into frames by the frame division means based on the voice division thus detected, and the second feature parameter extraction means for voice recognition is used from the divided frame data of a predetermined frame. Feature parameters are extracted, and the recognition unit recognizes the voice data using the extracted feature parameters.

Therefore, it is possible to accurately detect a voice signal with a small amount of calculation.

According to the fifth and sixth aspects of the invention, the signals input in time series are fetched as data and stored in the data storage means. Noise and a signal are detected by the signal detection means in the data stored in the data storage means, and the delimiter section of the signal detected by the frame division means is divided into frames.

Then, the characteristic parameter is extracted from the divided frame data of the predetermined frame by the characteristic parameter extracting means, and the recognizing means recognizes the data of the signal input in time series using the extracted characteristic parameter.

Therefore, the signals input in time series can be accurately recognized.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings.

1 to 5 are diagrams showing an embodiment of a voice recognition device and a signal recognition device according to the present invention, and this embodiment is an example applied to a voice recognition device for recognizing a uttered voice.

First, the structure will be described. FIG. 1 is a block diagram of the voice recognition apparatus 10, in which white arrows indicate the flow of data. In this figure, the voice recognition device 1
0 is an input unit 11 to which a voice signal or a music signal is input,
A sampling unit 12 that outputs a sound information pattern by sampling a voice signal or a music signal input to the input unit 11 at a predetermined sampling number, a memory unit 13 that stores the A / D converted voice signal, and will be described later. Audio signal detector 1
The feature parameter used for 5 is the feature extraction unit 1 described later.
6 (feature extraction unit 2), a feature extraction unit 14 (feature extraction unit 1) for extracting by a simple process that also functions as a pre-stage process, and a noise signal and voice that are configured by a neural network and are classified into several types in advance. The learning is performed by the signal, and the voice signal detection unit 15 that discriminates noise from a voice signal by neural network calculation based on the feature parameter sent from the feature extraction unit 14 (feature extraction unit 1), and the feature extraction unit 14 ( When the feature parameters of all the frames are sent from the feature extracting unit 1), the feature extracting unit 16 (feature extracting unit 2) that calculates the feature parameters for recognition by adding a predetermined calculation to the sent feature parameters ) And a feature extraction unit 16 (feature extraction unit 2), a recognition unit for performing a predetermined calculation based on the feature parameter to recognize a voice and output a recognition result. 7, a display unit 18 such as an LCD that displays the recognition result from the recognition unit 17, and a control unit 19 that receives signals from the above units and outputs a control signal to each unit to control the operation of each unit. ing.

The input section 11 includes a microphone 21 for inputting voice, an amplifier 22 for amplifying the input voice, and a low-pass filter 23 for removing noise of the input voice signal.
The input voice information and music information are converted into a predetermined analog signal and output to the sampling unit 12.

The sampling section 12 is composed of an A / D converter or the like, and samples and samples at a predetermined sampling interval (sampling frequency), and outputs the sampled data to the memory section 13.

The memory unit 13 is composed of a ring buffer structure RAM (Random Access Memory) or the like,
The audio signal converted into the digital signal according to the control signal from the control unit 19 is constantly updated and stored.

The feature extraction unit 14 (feature extraction unit 1) is
A feature parameter used in the audio signal detection unit 15 is extracted, and a relatively simple process that doubles as a preceding stage process of the feature extraction unit 16 (feature extraction unit 2) is performed. The feature extracting unit 14 (feature extracting unit 1) takes in data when the number of data stored in the memory unit 13 reaches a constant value (one frame), and extracts a feature parameter.

The voice signal detecting section 15 has a neural network structure as shown in FIG. 2, and is trained by noise signals and voice signals classified into several types in advance. In FIG. 2, the neural network 30 has a network structure composed of one layer of neurons (see circles) that should also be called as an interconnected associative layer and input synapses connected to all the associative layers, The voice signal detection unit 15 using the neural network 30 adds the feature parameters (Pk + 1 to Pl) sent from the feature extraction unit 14 (feature extraction unit 1) to the previously sent feature parameters (P0 to Pk). The characteristic parameters (P0 to Pl), and the characteristic parameters (P0
~ Pl) is used as an input and a predetermined operation is performed to generate a noise signal when O0 to Om of the outputs are ignited, and Om + 1 to
When On fires, it is determined that it is a voice signal and outputs the determination result.

FIG. 3 is a diagram showing a pre-stage process input to the input neuron of the above-mentioned neural network, and is an example using a two-dimensional mel-cepstrum transform as the pre-stage process.

The following processing is performed for each frame as the pre-stage processing input to the input neuron. That is, the input of the voice signal → Fourier transform → Mel transform (logarithmic conversion) → inverse Fourier transform → cepstrum is performed for each frame, and the cepstrum obtained by the above-described preprocessing is Fourier transformed for every 32 frames to obtain the result shown in FIG. Input to the input neuron of the neural network.

Next, the operation of this embodiment will be described.

By the way, as a voice recognition method for such a voice recognition device, an approach based on a neural network technique has been attempted.

The neural network is a parallel processing type in which a plurality of processors are closely coupled to each other as a model by using nerve cells of the brain as a model to solve a problem that the conventional sequential processing type computer system cannot handle. It is intended to be solved by a computer system, and exerts an excellent ability particularly in pattern-related problems such as pattern mapping, pattern perfection, and pattern recognition. Speech synthesis and speech recognition are examples of areas that contain many problems related to pattern recognition.
As the speech recognition method, various methods proposed in the neural network technology can be applied. When classified as a pattern classifier, various methods of neural networks can be classified according to whether an input signal takes a binary pattern or a continuous value, whether coupling efficiency is learned with a teacher or without a teacher, and the like. When the neural network method is applied to the voice recognition, since the voice changes continuously in time series, the method of taking a continuous value in the input pattern is applied. Among the neural networks that take continuous values, there are perceptrons and multilayer perceptrons as supervised ones, and there is a self-organizing feature map as unsupervised ones. Among these neural network methods, the one that realizes Voronoi division in the vector quantization method is a self-organizing feature map.

The self-organizing feature map is a neural network mapping method proposed by Kohonen of the University of Helsinki. Since the learning is performed without supervision, the network structure is simple and further It is composed of neurons that should be called interconnected layers, and input synapses that are connected to all connected layers. In the self-organizing feature map,
The input signals are sum-of-products calculated by the synaptic weight, and any specific neuron in the association layer is put into the output state. The specific neuron that is in the output state and the neighboring neurons are
The synaptic weights are modified to be most sensitive to similar input signals. At this time, it seems that the neuron with the highest output state first competes with other neurons for the same input signal and maximizes its own output, so this learning process is also called cooperative learning. There is. With this joint learning, when the activity of a specific neuron increases with respect to the input signal, the neuron in a close distance relationship is dragged to increase the activity, and the neurons outside the neuron have the opposite activity. The phenomenon that the value goes down occurs. The neurons whose activity is increased by this phenomenon are collectively called "bubbles". That is, the basic operation in the self-organizing feature map is that a certain "bubble" is generated on the two-dimensional plane of the neuron according to the input signal, and the "bubble" gradually grows and becomes a specific bubble. The synaptic weight of the neural network is automatically corrected so as to selectively respond to the input signal (stimulus). As a result of this operation, a specific input signal is finally associated with a "bubble", and the input signal is automatically classified.

In the present embodiment, by providing the voice signal detecting section 15 using the above neural network, the voice section is not detected simply by the voice level or the power level, but is learned in advance by the noise signal and the voice signal. Since the voice section is detected by the neural network, it becomes possible to accurately detect the voice signal.

FIG. 4 is a flow chart showing the voice recognition processing of the voice recognition apparatus 10, and in the figure, reference numeral Sn (n = n =
1, 2, ...) Indicate each step of the flow.

First, in step S1, one piece of voice data is fetched, the fetched voice data is written in the memory section 13, and in step S2 it is judged whether or not the data stored in the memory section 13 is one frame. To do. Next, in step S3, the feature extraction unit 14 (feature extraction unit 1) sets 1
The data for the frame is fetched to extract the characteristic parameter, and in step S4, the speech signal detecting section 15 calculates the neural network for detecting the speech signal by using the characteristic parameter sent from the characteristic extracting section 14 (feature extracting section 1) as an input. To do. That is, in the voice signal detection unit 15, the feature extraction unit 1
4 (feature extraction unit 1) sends the feature parameter (Pk +
1 to Pl) is added to the previously sent characteristic parameter (P0 to Pk) to make it a characteristic parameter (P0 to Pl), and this characteristic parameter (P0 to Pl) is used as an input to perform a predetermined calculation and output When O0-Om ignites, a noise signal
Further, when Om + 1 to On are fired, a voice signal detecting neural network calculation for determining a voice signal is performed.

Next, in step S5, as a result of the neural network calculation, it is determined whether or not the signal from the voice signal detecting section 15 is a voice signal. If the signal from the voice signal detecting section 15 is not a voice signal, a noise signal is used. If it is determined that the signal is present, it is determined in step S6 whether or not the previous signal from the audio signal detection unit 15 is an audio signal. If the previous signal from the voice signal detecting unit 15 is not a voice signal, it is determined that it is a noise signal, and in step S7 the address of the head of the frame is shifted by 1/2 frame and the process returns to step S1 to fetch one data. . That is, the feature extraction unit 14 (feature extraction unit 1)
Shifts the address (pointer) at the head of the frame by 1/2 frame, returns to step S1, waits for the data for one frame to be stored, and extracts the characteristic parameter.
The voice signal detection unit 15 performs the neural network calculation while updating the feature parameter from the feature extraction unit 14 (feature extraction unit 1). The processes of steps S1 to S7 are repeated until a voice signal is detected in step S5 or step S6. When the voice signal is detected in step S5, whether or not the voice signal is initially detected is detected in step S8. When the voice signal is initially detected, the word address of the voice signal detected in step S9 is set. Then
In step S10, it is determined whether the memory is over, and if the memory is not over, it is determined that the audio signal from the audio signal detector 15 has not reached a predetermined length, and the data is stored in the memory until the memory is over. In order to write, the process proceeds to step S7 and the above steps S1 to S7.
The process of is repeated.

When the memory is over in step S10 or when the previous voice signal is detected in step S6, it is determined that the voice signal from the voice signal detecting section 15 has reached a predetermined length, and step In step S11, the ending address is set, in step S12 the word length (speech section) is calculated based on the result of the beginning and ending judgment, and the frame of the speech section is divided. In step S13, the feature extraction unit 16 (feature extraction unit 2 ), The characteristic parameter of one frame is extracted. Next, in step S14, it is determined whether or not the extraction of the characteristic parameters has been completed for all the frames. If the extraction of the characteristic parameters of all the frames has not been completed, the process returns to step S13 to extract the characteristic parameters of all the frames. The extraction of the characteristic parameter is repeated until. When the extraction of the feature parameters of all frames is completed, the recognition unit 17 causes the feature extraction unit 16 to perform the extraction in step S15.
The feature neural network for recognition is calculated using the feature parameters of all the frames sent from the (feature extraction unit 2), and the recognition result is displayed on the display unit 18 in step S16, and the processing of this flow ends.

By executing the above processing, specifically, the following voice recognition operation is performed in each unit of the voice recognition apparatus 10.

After the voice input from the microphone 21 is amplified by the amplifier 22, the noise of the voice signal input by the low-pass filter 23 is removed, converted into a digital signal by the A / D converter 12, and the control unit 19 In accordance with the control signal from the above, it is constantly updated and stored in the memory unit 13 of the ring buffer structure.

The feature extraction unit 14 (feature extraction unit 1)
A feature parameter used in the audio signal detection unit 15 is extracted, and a relatively simple process that doubles as a preceding stage process of the feature extraction unit 16 (feature extraction unit 2) is performed. In the feature extraction unit 14 (feature extraction unit 1), when the number of data stored in the memory unit 13 reaches a fixed value (for one frame), the data is taken in to extract the feature parameter and the voice signal detection unit 15 is extracted. send.

The voice signal detection unit 15 is composed of a neural network as shown in FIG. 2, and is trained by noise signals and voice signals that are classified into several types in advance. The voice signal detection unit 15 using the neural network 30 adds the feature parameters (Pk + 1 to Pl) sent from the feature extraction unit 14 (feature extraction unit 1) to the previously sent feature parameters (P0 to Pk). Then, the characteristic parameters (P0 to Pl) are set, and a predetermined calculation is performed using the characteristic parameters (P0 to Pl) as an input to output O0 to Om of the outputs.
When it is fired, it is determined to be a noise signal, and when Om + 1 to On are fired, it is determined to be a voice signal and the determination result is output to the control unit 19.

When the signal from the audio signal detection unit 15 is determined to be a noise signal, the control unit 19 causes the feature extraction unit 14
The (feature extraction unit 1) is instructed to continue the audio signal state.

Then, the feature extraction unit 14 (feature extraction unit 1) shifts the data points by ½ frame and waits for one frame of data to be stored again, and extracts the feature parameter to extract the voice signal detection unit. Send to 15. The voice signal detection unit 15 performs calculation while updating the characteristic parameter.

The above-described operations (1) to (5) are performed by the audio signal detecting section 15
Repeat until the voice signal is detected by.

When the audio signal detection unit 15 detects an audio signal, the control unit 19 writes the data to the memory while the audio signal detection signal is being received from the audio signal detection unit 15 or after writing data in the memory for a predetermined time. Prohibit writing of data.

After performing the frame division of the voice section, the control unit 19 sends the start point of each frame to the feature extraction unit 14 (feature extraction unit 1) so that the feature extraction is performed again for each frame. The feature extraction unit 16
(Feature extraction unit 2).

When the parameters of all the frames are sent from the feature extraction unit 14 (feature extraction unit 1), the feature extraction unit 16 (feature extraction unit 2) performs calculation on these to calculate the recognition parameters and recognizes them. Send to section 17.

The recognition unit 17 performs calculation based on the characteristic parameters sent from the characteristic extraction unit 16 (feature extraction unit 2) and sends the recognition result to the display unit 18, and the display unit 18
Displays the voice recognition result from the recognition unit 17 according to the instruction of the control unit 19 and ends the series of voice recognition operations.

The control unit 19 outputs a control signal to each unit to return the operation state of the voice recognition apparatus 10 to the above state and start the next recognition.

As described above, the voice recognition device 10 of the present embodiment is composed of the feature extraction unit 14 (feature extraction unit 1) for extracting the feature parameters for voice signal detection and the neural network, and has various types in advance. Learning is performed by the noise signal and the voice signal classified into
A voice signal detection unit 15 that discriminates noise from a voice signal by neural network calculation based on the feature parameters sent from the feature extraction unit 14 (feature extraction unit 1), and the feature extraction unit 14
When the feature parameters of all the frames are sent from the (feature extracting unit 1), the feature extracting unit 16 (feature extracting unit) that calculates a feature parameter for recognition by performing a predetermined calculation on the sent feature parameters 2) and the feature extraction unit 1
6 (feature extracting unit 2), a recognition unit 17 that performs a predetermined calculation based on the feature parameters sent from the feature recognition unit to recognize a voice and output a recognition result, and a display unit 18 that displays the recognition result from the recognition unit 17. A control section 19 for outputting a control signal to control the operation of each section is provided, and a voice signal detecting section 15 using a neural network detects a voice section by a neural network previously learned from a noise signal and a voice signal. As a result, a voice signal can be accurately detected without interruption in real time, and a voice recognition device that is resistant to noise and changes can be realized. That is, in the conventional voice recognition device, since the detection of the voice section is simply performed by the voice level or the power level, the voice section cannot be detected or detected when the voice level is low or the noise level is high. However, the voice recognition apparatus of this embodiment is provided with the voice signal detection unit 15 using a neural network, so that the noise is simply noise to the signal (S / N ratio). The noise component is also input as a noise signal to the signal as the input of the neural network, and the voice section is detected by the neural network calculation of the noise signal and the voice signal, so it is not affected by the magnitude of the signal level. It becomes possible to detect an audio signal accurately.

Further, the feature extraction unit 14 for calculating the feature parameter for voice signal detection used by the voice signal detection unit 15
Since the (feature extraction unit 1) is a part of the feature parameter extraction originally used for the recognition, the feature extraction unit 14 can be realized without newly adding hardware or software.

Further, in this embodiment, since the recognition unit 17 uses the neural network, it is possible to obtain the accurate voice recognition result in real time.

Although the present embodiment is an example applied to a voice recognition device for recognizing a uttered voice, of course, the present invention is not limited to this, and a voice signal detecting section for detecting a noise signal and a voice signal of voice data. It goes without saying that any device provided with can be applied to all devices. For example, a device that performs similar recognition processing on a signal input in time series (including a voice signal, but not limited to a voice signal) may be used.

Further, in this embodiment, the feature extracting unit 14 (feature extracting unit 1) is provided, and the feature extracting unit 16 (feature extracting unit 2) is provided.
The characteristic parameters for audio signal detection are extracted by a simple process that doubles as the preceding stage processing, but if the data can be supplied to the audio signal detection unit that distinguishes between a noise signal and an audio signal, this characteristic The extraction unit 14 may have any configuration, and may not have such a feature extraction unit.

Further, in the present embodiment, the voice signal detecting section 15
By using a neural network for the recognition unit 17 and the recognition unit 17, it is possible to obtain an accurate speech recognition result. However, any form of this neural network may be used, and the recognition unit 17 does not use a neural network. It goes without saying that it may be one.

Further, it is needless to say that the number and types of the circuits and members constituting the above speech recognition apparatus are not limited to those in the above-mentioned embodiment, but may be realized by software.

[0056]

According to the first, second, third, fourth, fifth and sixth aspects of the present invention, since the voice signal detecting means for detecting the noise signal and the voice signal of the voice data is provided, the calculation amount is small. In addition, it is possible to realize a voice recognition device that can detect a voice signal accurately and accurately and that is resistant to noise and changes.

According to the invention described in claims 7, 8 and 9,
Since the signal input in time series is captured as data and the noise signal and voice signal of this data are detected, the amount of calculation is small and the time series signal can be detected accurately with respect to the signal input in time series. A signal recognition device that is resistant to noise and changes can be realized.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of a voice recognition device.

FIG. 2 is a diagram showing a neural network structure of a voice signal detection unit of the voice recognition device.

FIG. 3 is a diagram showing a pre-stage process of a neural network of the voice recognition device.

FIG. 4 is a flowchart showing a voice recognition process of the voice recognition device.

FIG. 5 is a diagram showing overlapping of frames of the voice recognition device.

[Explanation of symbols]

 10 voice recognition device 11 input unit 12 sampling unit (A / D converter) 13 memory unit 14 feature extraction unit (feature extraction unit 1) 15 voice signal detection unit 16 feature extraction unit (feature extraction unit 2) 17 recognition unit 18 display unit 19 control unit 21 microphone 22 amplifier 23 low-pass filter 30 neural network

Claims (6)

[Claims] 1. A voice data storage unit for storing voice data; a voice signal detection unit for detecting a noise signal and a voice signal of the voice data stored in the voice data storage unit; and a voice signal detection unit for detecting the noise signal and the voice signal. A frame dividing unit that divides the voice section of the audio signal into frames; a characteristic parameter extracting unit that extracts a characteristic parameter from the frame data divided by the frame dividing unit; and a characteristic parameter that is extracted by the characteristic parameter extracting unit. A voice recognition device comprising: a recognition unit that recognizes voice data by using the recognition unit. 2. A voice data storage unit for storing voice data, a first feature parameter extraction unit for extracting a feature parameter used by a voice signal detection unit from the voice data stored in the voice data storage unit, A voice signal detection unit that detects a noise signal and a voice signal based on the feature parameter output from the first feature parameter extraction unit, and a frame that divides the voice section of the voice signal detected by the voice signal detection unit into frames. By using the dividing means, the second characteristic parameter extracting means for extracting the characteristic parameter for voice recognition from the frame data divided by the frame dividing means, and the characteristic parameter extracted by the second characteristic parameter extracting means. A voice recognition device comprising: a recognition unit for recognizing voice data. 3. The voice recognition device according to claim 1, wherein the voice signal detection means is configured by a neural network. 4. The voice signal detecting means is composed of a neural network, and learning is performed by a noise signal and a voice signal which are classified into a predetermined type in advance, and the noise signal and the voice signal are discriminated by a neural network calculation. The voice recognition device according to claim 1 or 2, wherein 5. A data storage means for storing signals inputted in time series as data, a signal detection means for detecting noise and a signal of the data stored in the data storage means, and a signal detection means for detecting the noise and the signal. A frame dividing unit that divides a signal delimiter section into frames, a characteristic parameter extracting unit that extracts a characteristic parameter from the frame data divided by the frame dividing unit, and a characteristic parameter that is extracted by the characteristic parameter extracting unit are used. A signal recognition device comprising: a recognition unit that recognizes time series data. 6. The signal recognition device according to claim 5, wherein the voice signal detection means is configured by a neural network.