JPH06149290A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To improve the recognition rate of the speech recognizing device in noisy environment by absorbing the distortion of a feature parameter inputted to a microphone for speech input without being affected by the position of a noise source in a cabin and the position of the microphone for speech input, and efficiently performing noise removal. CONSTITUTION:The speech recognizing device is equipped with the microphone 1 to which a speech and an acoustic noise are inputted indirectly through a space, a sound signal input terminal 3 to which they are inputted directly from an acoustic device, an acoustic characteristic correcting means (sound field correction part 8) which compensates characteristics of the direct signals, a 1st feature extraction part 2 which extracts a 1st feature parameter by analyzing the input signal from the microphone 1, a 2nd feature extraction part 4 which extracts a 2nd feature parameter by analyzing the corrected input signal from the sound field correction part 8, a noise removal part 5 which generates a 3rd feature parameter by subtracting the 2nd feature parameter from the 1st feature parameter, a speech pattern generation part 6 which generates a speech pattern from the 3rd feature parameter, and a recognizing process part 7 which recognizes the speech pattern.

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,
In particular, the present invention relates to a voice recognition device that eliminates the influence of sound (which becomes noise for voice recognition) generated by an audio device such as a television or a radio.

[0002]

2. Description of the Related Art In the case of performing voice recognition, first, a speaker utters a voice into a microphone and inputs this voice into the microphone. At this time, the utterance environment of the speaker is not always good. It is not limited, but rather, some kind of noise is generally present. Therefore, in the case of voice input, noise is usually superimposed and input in many cases, and appropriately eliminating the influence of noise leads to realizing voice recognition with a high recognition rate.

Therefore, in recent years, various methods for removing such noise have been proposed. Examples of such conventional noise removal methods include, for example, Japanese Patent Application Publication No. Showa 54-91007.
A voice recognition device described in Japanese Patent Publication No.
The noise signal N input from the second microphone is subtracted from the noise-superimposed voice signal N + V (voice signal V + noise signal N) input from the first microphone using the second microphone for noise input and the second microphone for noise input. Then, there is a case where an audio signal V without noise is tried to be obtained.

In such a conventional technique, it is difficult to set the installation position of the noise input microphone. For example, when the noise input microphone is installed near the voice input microphone, the noise is input to the noise input microphone. There is an inconvenience that the input is relatively large. On the contrary, when the voice input microphone and the noise input microphone are installed far apart from each other, the noise input with the voice input microphone and the noise input to the noise input microphone The characteristics (phase, frequency, volume, etc.) do not match, and even if these differences are taken, there is the inconvenience that effective noise removal cannot be achieved.

Thus, depending on the environment in which voice recognition is performed,
For example, in an environment where there is sound such as music or voice emitted from the speakers of audio equipment such as TV and radio,
These sounds become noise for voice recognition. In such a television or radio with a clear noise source, the acoustic signals generated by them can be detected directly from the output terminals of these acoustic devices, so that the noise input microphone as described above can be omitted.

However, even when the acoustic signal N such as music or voice generated by such an acoustic device is directly obtained from these devices, the direct acoustic signal N and the voice are input to the voice input microphone for voice recognition together with the voice. The input indirect acoustic voice Nz is often characteristically different. For example, in the interior of an automobile, the music (acoustic noise) emitted from the car stereo speakers is affected by the acoustic characteristics of the space inside the automobile, so the signal characteristics such as phase, frequency, and volume change. The resulting signal Nz is different from that of the signal N of music (acoustic noise) obtained from the output terminal of the car stereo, and there is a fear that the signal Nz is different in characteristics.

Therefore, as described above, the signal characteristic of the direct acoustic signal N from the output terminal of the audio equipment and the signal characteristic of the indirect acoustic signal Nz, which is the sound emitted from the speaker and distorted by the influence of the spatial environment, is detected by the microphone. , The difference signal V + Nz-N is true even if the direct acoustic signal N from the output terminal of the audio device is subtracted from the combined signal Nz + V of the indirect acoustic signal Nz and the audio signal V input to the microphone. Therefore, it is impossible to realize voice recognition with a high recognition rate after all by such noise removal processing.

[0008]

DISCLOSURE OF THE INVENTION The present invention has been made in view of the above-mentioned points, and the signal characteristics of the direct acoustic signal N from the output terminal of the acoustic device and the spatial characteristics of the spatial environment emitted from the speaker. When the acoustic characteristics are distorted due to the influence and the acoustic characteristics are different from the indirect acoustic signal Nz detected by the microphone, the difference in these acoustic characteristics (phase, frequency, volume, etc.) should be corrected (N = Nz). By this, a true voice signal V is obtained, and a highly accurate voice recognition device is realized.

[0009]

A voice recognition apparatus of the present invention includes a microphone to which a voice uttered for voice recognition is input and the acoustic noise is indirectly input via the space. An acoustic signal input terminal in which an acoustic noise signal is directly input from the acoustic device without passing through the space, an indirect signal of acoustic noise obtained from the microphone, and a direct signal of acoustic noise obtained from the acoustic signal input terminal. Acoustic characteristic correction means for characteristic-correcting a direct signal of acoustic noise input from the acoustic signal input terminal, and an input signal from the microphone are analyzed to reduce a characteristic difference between First characteristic extracting means for extracting a parameter, second characteristic extracting means for analyzing a corrected input signal from the acoustic characteristic correcting means to extract a second characteristic parameter, and the first characteristic parameter The noise removal means for subtracting the second characteristic parameter from the data to generate the third characteristic parameter, the voice pattern generating means for generating a voice pattern from the third characteristic parameter, and the voice pattern generating means. And a recognition processing means for performing recognition processing on the voice pattern.

[0010]

In the voice recognition apparatus of the present invention, the direct acoustic signal N from the output terminal of the acoustic device is corrected by the acoustic characteristic correction means.
Of the indirect acoustic signal Nz detected by the microphone. That is, the direct acoustic signal N from the output terminal of the audio device is also subjected to characteristic distortion due to the influence of the spatial environment to have characteristics (phase, frequency, volume, etc.) that are not equal to the indirect acoustic signal Nz. Thereby, the signal characteristic of the corrected direct acoustic signal Nz obtained by correcting the direct acoustic signal N of the output terminal of the acoustic device from the signal characteristic of the combined signal Nz + V of the indirect acoustic signal Nz and the audio signal V detected by the microphone. True audio signal V minus
Based on the, voice recognition can be performed.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment in which the speech recognition processing apparatus according to the present invention is used in the interior of an automobile equipped with an acoustic device will be described below.

FIG. 1 shows an example of the configuration of a voice recognition processing device according to the present invention. In FIG. 1, 1 is a microphone for inputting a voice, and 2 is a first feature extraction unit for analyzing the input voice and extracting a first feature parameter. Reference numeral 3 is a noise input terminal for directly inputting an output (acoustic noise) from an output terminal of the audio device, and when the audio device has four speakers, there are four noise input terminals.

Reference numeral 4 is a second feature extraction unit, which corrects a direct signal obtained from the noise input terminal 3 by a sound field correction unit 8 which will be described later, analyzes the noise, and outputs a second feature parameter thereof. To extract.

The first and second feature extracting units 2 and 4
For the same, the same analysis process, for example, an FFT processing device is used.

Reference numeral 5 is a noise removing unit for extracting the third characteristic parameter by subtracting the second characteristic parameter from the first characteristic parameter.

Reference numeral 6 denotes a pattern creating section for creating a voice pattern from the third characteristic parameter obtained from the noise removing section 5. Reference numeral 7 denotes a recognition processing unit that performs recognition processing of a voice pattern obtained from the pattern creation unit 6, and a plurality of voice patterns to be recognized have already been registered. As a result of comparison with an unknown voice pattern, the most similar voice pattern is obtained. By identifying a registered voice pattern, the voice is identified, that is, recognized.

Reference numeral 9 is a stationary noise characteristic parameter storage unit for storing characteristic parameters of stationary noise (running noise such as engine sound in the vehicle compartment) as necessary. Used for noise reduction processing. That is,
When the third feature parameter is extracted by subtracting the second feature parameter from the first feature parameter, the stationary noise feature parameter of the storage unit 9 is also subtracted to obtain the third feature parameter.
The characteristic parameters of

Here, the sound field correction unit, which is the most characteristic feature of the voice recognition apparatus of the present invention, will be described in detail. This sound field correction unit 8
Is to reduce the difference in characteristics (phase, frequency, volume, etc.) between the indirect signal of acoustic noise obtained from the microphone 1 and the direct signal of acoustic noise obtained from the noise input terminal 3. It is an acoustic characteristic correction means for correcting the direct signal of the acoustic noise input from the input terminal 3.

FIG. 2 shows the configuration of the sound field correction section 8 as such an acoustic characteristic correction means. In the figure, 80, 80
.. is an A / D converter that converts an input acoustic noise signal from an analog signal to a digital signal.

Reference numeral 81 is a gain correction data table that holds gain correction data, 82 is a gain correction unit that corrects the gain using this gain correction data, and 83 is frequency characteristic correction data that holds frequency characteristic correction data. A data table, 84 is a frequency characteristic correction unit that corrects frequency characteristics using the frequency characteristic correction data, and 85 is a phase correction data table that holds phase correction data.
Is a phase correction unit that corrects the phase using the phase correction data.

Further, 87 is an adder for adding each of the four outputs from the phase corrector 86 for all four speakers, and 88 is a D / A converter for converting the digital signal of the added value into an analog signal. is there.

In the sound field correction unit 8 having such a configuration, the voice microphone 1 and the noise input terminal 3 are provided before performing voice recognition.
The difference in the characteristics of the acoustic data that occurs between and is measured, and the characteristic correction data is created.

Specifically, for example, a pulse sound is generated in the audio device, and the pulse sound output from the speaker and influenced by the spatial environment such as the interior of the automobile is input to the microphone 1.
The noise input terminal 3 is directly input from the output terminal of the audio device in the form of an electric signal. At this time, the gain correction unit 82 adjusts the sound volume, the frequency characteristic correction unit 84 adjusts the frequency characteristic, and the phase correction unit 86 adjusts the phase. As a result of each of these adjustments, the noise removal unit 5 causes the microphone 1 to operate. Gain correction data table 81, frequency characteristic correction data table 83, so that the difference between the characteristic parameters (third characteristic parameter) obtained from both the noise input terminal 3 and the noise input terminal 3 is minimized.
The respective values of the phase correction data table 85 are set.

That is, the gain correction unit 82, the frequency characteristic correction unit 84, and the frequency characteristic correction unit 84 perform correction adjustment of each parameter while feeding back the difference result of the noise removal unit 5, and the third characteristic parameter exists. When it is not performed or when it becomes the minimum, the correction data (constant or an appropriate function) of each parameter is stored in the gain correction data table 81, the frequency characteristic correction data table 83, and the phase correction data table 85.

The recognition processing operation will be described below with reference to FIGS. 1 and 2.

First, by the first feature extraction unit 2,
The voice analysis of the voice input from the microphone 1 is performed.
In the first feature extraction unit 2, the first feature parameter is extracted by converting the input analog audio signal into a digital signal and then performing a Fourier transform on the obtained digital signal. The first characteristic parameter obtained here is further sent to the noise removing unit 5.

While voice is being input from the microphone 1,
The acoustic noise of the car stereo is directly input to the noise input unit 3 from its output terminal. For example, when there are four speakers, the acoustic noise input to the noise input terminal 3 is
This is done via four input terminals.

The four systems of acoustic noise signals input to these four input terminals 3 are sent to the sound field correction section 8. The sound field correction unit 8 corrects the difference in phase, frequency characteristic, and volume generated between the acoustic noise signal and the acoustic noise signal input from the microphone 1 for each of these four systems.

That is, in the sound field correction unit 8, as shown in FIG. 2, the A / D converter 80 first converts each input noise signal into a digital value.

Then, the gain correction section 82 uses the gain correction data recorded in the gain correction data table 81 to correct the gain for each of the four outputs from the A / D converter 80.

Next, the frequency characteristic correction unit 84 uses the frequency characteristic correction data stored in the frequency characteristic correction data table 83 to obtain each 4 from the gain correction unit 82.
The frequency characteristic is corrected for the output.

Subsequently, the phase correction unit 86 uses the phase correction data recorded in the phase correction data table 85 to correct the phase for each of the four outputs from the frequency characteristic correction unit 84.

When these corrections are completed, the adding section 87
However, each of the four outputs from the above-mentioned final stage phase correction unit 86 is added to one acoustic noise signal, and the D / A converter 8
8 converts this into an analog signal.

This is the end of the processing of the sound field correction unit 8.
The acoustic noise signal corrected by the sound field correction unit 8 as described above is processed by the second feature extraction unit 4 and the above-described first feature extraction unit 2
Then, the same voice analysis process is performed to extract the second characteristic parameter.

The first characteristic parameter created from the input of the microphone 1 and the second characteristic parameter created from the input of the noise input unit 3
The characteristic parameters of are transmitted to the noise removing unit 5.

The noise removing section 5 basically subtracts the second characteristic parameter consisting of acoustic noise from the first characteristic parameter consisting of acoustic noise and voice to generate a third characteristic parameter not containing acoustic noise. .

Further, the pattern creating section 6 detects the voice section of the third characteristic parameter and creates a voice pattern.

The recognition processing section 7 performs recognition processing by comparing the voice pattern thus created with the voice pattern registered in the recognition processing section 7.

As described above, according to the present invention, the defect caused by the position of the noise source, the position of the voice input microphone, etc., that is, the characteristic parameter of the actual noise itself and the actual noise propagate in the space. It becomes possible to correct the difference in the characteristics (phase, frequency, volume, etc.) that occur between the characteristic parameters of the noise input from the microphone after being affected by this spatial environment during operation. In an environment where acoustic noise is constantly changing, such as in a car equipped with a stereo, noise is reliably removed, and deterioration of recognition performance due to noise is prevented.

In the present embodiment, the case where the voice recognition device according to the present invention is used in the interior space of a car equipped with a car stereo has been described, but the present invention also applies to the acoustic data of noise in other spaces. Can be applied in any environment where is known by another means before is input to the speech recognition device via the speech recognition space.

Further, in order to improve the recognition accuracy in the voice recognition processing apparatus which operates in the above-mentioned environment, noise (steady noise) such as engine sound, running sound, and sound generated by natural phenomenon is often taken into consideration. It is necessary to. Hereinafter, a case where such stationary noise is removed in the voice recognition processing device will be described.

The stationary noise is basically removed by subtracting the stationary noise characteristic parameter of the stationary noise parameter storage unit 9 from the obtained third characteristic parameter in the noise removing unit 5 according to the change. Done. Here, the stationary noise parameter is a first characteristic parameter created in a state in which there is no acoustic noise input to the noise input terminal 3 and no voice or acoustic noise input to the microphone 1. It means that.

Specifically, since the audio device is used as a noise source in this embodiment, the first characteristic parameter is extracted in advance in a state where nothing is output from the audio device, and this is used as stationary noise. It may be stored in the parameter storage unit 9.

Therefore, at the time of recognition, the input from the microphone 1 and the input from the noise input terminal 3 are processed by the method as described above, and the first characteristic parameter and the second characteristic parameter are extracted.

The noise removing section 5 subtracts the second characteristic parameter from the first characteristic parameter and further subtracts the stationary noise parameter of the stationary noise parameter storage section 9 to create a third characteristic parameter. The pattern creating section 6
Third, the stationary noise parameters are subtracted as described above
The voice pattern is created by detecting the voice section with respect to the characteristic parameter of. The recognition processing unit 7 performs recognition processing by comparing the voice pattern created in this way with the voice pattern registered in the recognition processing unit 7.

Further, since the stationary noise generated by the engine noise or the like may change gently, in this case, in order to effectively remove the stationary noise, the stationary noise parameter of the stationary noise parameter storage unit 9 is required. It is preferable to update according to. Such update processing will be described next.

First, if the value of the third characteristic parameter is less than this, it can be concluded that ambient noise and voice from the car stereo have not been input to the voice recognition processing device (input is only stationary noise). Set a threshold like this.

Therefore, the noise removing unit 5 checks the value of the third characteristic parameter at every constant period (for example, 10 msec) at the time of recognition, and when the value of the third parameter is equal to or less than the threshold value, the stationary noise is detected. Is assumed to have been input from the microphone, and the stationary noise parameters are updated without detecting the voice input.

In such an updating method, for example, the result of weighted averaging the stationary noise parameter and the third characteristic parameter at a ratio such that the ratio of the stationary noise parameter is high is registered as a new stationary noise parameter.

In practice, the stationary noise parameter ratio is increased (for example, at a ratio of 15: 1) to perform weighted averaging so that the stationary noise parameter value does not change significantly.

As described above, according to the present invention, stationary noise such as engine noise in an automobile can be effectively dealt with, so that the deterioration of recognition performance due to noise can be prevented.

In this embodiment, the frequency characteristic is changed by FFT or IFFT, but it may be changed by other means. It is also possible to perform sound field correction after feature extraction. Also, in this example,
The input to the noise input section was directly input from the speaker,
It is also possible to install and process a microphone so that desired voice is not input. These processes can be implemented by a circuit and a DSP.

[0053]

According to the speech recognition apparatus of the present invention, the direct acoustic signal N from the output terminal of the acoustic device is also subjected to the characteristic distortion due to the influence of the spatial environment, and the indirect acoustic signal N of the microphone input.
Since characteristics (phase, frequency, volume, etc.) that are not equal to z can be obtained, the indirect acoustic signal Nz and the voice signal V detected by the microphone are generated.
The voice recognition can be performed based on the true voice signal V obtained by subtracting the signal characteristic of the corrected direct acoustic signal Nz obtained by correcting the direct acoustic signal N of the output terminal of the acoustic device from the signal characteristic of the combined signal Nz + V of Therefore, it is possible to realize voice recognition with a high recognition rate even in a noisy environment.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of a voice recognition processing device of the present invention.

FIG. 2 is a block configuration diagram of a sound field correction means 8 of the device of the present invention.

[Explanation of symbols]

 1 Microphone 2 1st feature extraction part 3 Noise input terminal 4 2nd feature extraction part 5 Noise removal part 6 Pattern creation part 7 Recognition processing part 8 Sound field correction part 82 Gain correction part 84 Frequency characteristic correction part 86 Phase correction part

Claims (1)

[Claims] 1. In a voice recognition device for performing voice recognition in a spatial environment in which the sound from the audio device becomes noise for voice recognition, a voice uttered for voice recognition is input and the acoustic noise is generated. A microphone that is indirectly input via the space, an acoustic signal input terminal to which the acoustic noise signal is directly input from the audio device without passing through the space, and an indirect acoustic noise obtained from the microphone. Acoustic characteristic correcting means for correcting the characteristic of the direct signal of the acoustic noise input from the acoustic signal input terminal so as to reduce the characteristic difference between the signal and the direct signal of the acoustic noise obtained from the acoustic signal input terminal; First feature extracting means for analyzing the input signal from the microphone to extract a first feature parameter, and second for analyzing the corrected input signal from the acoustic characteristic correcting means.
Second feature extracting means for extracting the feature parameter of No. 3, noise removing means for subtracting the second feature parameter from the first feature parameter to generate a third feature parameter, and a voice pattern from the third feature parameter. A voice recognition apparatus comprising: a voice pattern creating means for creating a voice pattern; and a recognition processing means for performing a recognition process on the voice pattern created by the voice pattern creating means.