JPH11327593A - Voice recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve voice recognition rate by correcting distortion generated at the time of suppressing noise included in an input signal mixedly including a voice to be recognized and the noise by using a spectrum subtraction method to an optimum level and reducing the distortion as much as possible even in the case of using the voice recognition system under an environment generating each different noise sort. SOLUTION: In a noise suppressing device 10, a mapping calculation part 17 constituted by a neutral network model corrects the distortion of a noise subtraction power spectrum outputted from a power spectrum clipping calculation part 16. When a voice is correctly recognized by the voice recognition device 20, a teacher power spectrum previously stored in a buffer 26a built in a judgement part 26 is outputted to a learning control part 18 in the device 10. The control part 18 uses a noise subtraction spectrum stored in a buffer 17a built in the calculation part 17 as an input and uses a teacher spectrum outputted from the judgement part 16 as a teacher output to control a parameter of a neural network.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to noise suppression used as pre-processing of speech signal processing such as speech recognition.
The present invention relates to a speech recognition system using a noise suppression device using a spectrum subtraction method for removing a noise component from an input signal in which a speech signal to be recognized and a noise signal are mixed as much as possible.

[0002]

2. Description of the Related Art Conventionally, a speech recognition device effective for, for example, a case where a destination setting in a car navigation system can be input by voice has been proposed and realized. In such a speech recognition device, an input speech is compared with a plurality of candidate patterns for comparison stored in advance, and a speech with a high degree of coincidence is used as a recognition result. It may not be completely accurate. This is true even in a quiet environment, especially in an environment where noise is generated in the surroundings. In particular, it is difficult to assume that there is no noise in consideration of an actual use environment such as the car navigation system described above. Therefore, in order to improve the recognition rate, it is necessary to perform noise suppression as much as possible to remove noise components from an input signal in which a speech signal and a noise signal necessary for recognition are mixed, as a preprocessing of an input to a speech recognition device. Is desirable.

As a method of removing a noise component from an input signal in which speech and noise are mixed, a spectrum subtraction method is known as a very effective method.
Regarding this spectrum subtraction method, for example, STEVEN F BOLL, “Suppression of Acoustic Noise i
n Speech Using Spectral Subtruction ”, IEEE Transa
ctions on Acoustics, Speech and Signal processin
g, Vol.Assp-27, No.2, April 1979, pp.113-120, and many other research results have been published. The spectrum subtraction method realizes noise suppression by subtracting the amplitude spectrum of noise from the amplitude spectrum of an audio signal containing noise or by subtracting the power spectrum of noise from the power spectrum of an audio signal containing noise. . The power spectrum is obtained by squaring the amplitude spectrum. The output by the spectrum subtraction method is an amplitude spectrum in which noise is suppressed or a power spectrum in which noise is suppressed.

As a system configuration for performing speech recognition after performing such noise suppression, for example, FIG.
A speech recognition system 200 as shown in FIG. That is, only a voice signal or a noise signal mixed with noise is input from the microphone 201. An input signal from the microphone 201 is input to the noise suppression device 203, and the speech signal noise-suppressed by the noise suppression device 203 is transferred to the speech recognition device 204. Also, in this case, the PTT
(Push-To-Talk) While pressing switch 205, microphone 2
A voice is input through the input unit 01. Then, noise suppression in the noise suppression device 203 is performed as follows.

[0005] That is, as shown in FIG.
The noise suppression device 203 takes in an input signal from the microphone 201, assuming that a noise section is present until the switch 205 is pressed. Then, when the PTT switch 205 is pressed, it is determined that the input signal is in the voice section, and the noise suppression device 203 takes in the input signal from the microphone 201. However, what is captured in the voice section is “voice signal + noise signal”. Therefore, by subtracting the “noise signal” captured in the noise section from the “voice signal + noise signal” captured in the voice section, a voice signal in which the noise signal is suppressed can be extracted.

[0006]

However, this technique is basically based on estimated noise. That is, in the voice section shown in FIG. 2B, the noise mixed is not directly detected, but the noise in the voice section is estimated based on the noise signal captured in the noise section before the start of the voice section, The power spectrum of the estimated noise is subtracted from the power spectrum of the input voice captured in the voice section. And, in general,
A value obtained by multiplying the power spectrum of the estimated noise by a predetermined coefficient (subtract coefficient) is subtracted from the power spectrum of the input voice, and the subtract coefficient is often set to a value larger than 1. As described above, setting the subtract coefficient to a value larger than 1 corresponds to subtracting more than necessary when subtracting the power spectrum of the estimated noise.

In a section in which the power of the voice is secured to some extent, such as a vowel part of the voice, even if the power spectrum of the estimated noise is slightly reduced, the shape of the power spectrum of the voice is hardly affected. However, when the power of the sound is low, such as in a pause section or a fricative consonant part in the sound, the sound may be overdrawn and become a negative value. As described above, since the power spectrum is obtained by squaring the amplitude spectrum, it cannot theoretically be a negative value. For this reason, a portion which is too negative and becomes a negative value is clipped and set to zero (0) or a relatively small positive constant. Therefore, the power spectrum of the noise-suppressed input voice obtained by the spectrum subtraction method has a specific distortion.

[0008] The same applies to the case where the amplitude spectrum is used without using the power spectrum. In other words, subtracting the value obtained by multiplying the amplitude spectrum of the estimated noise by the subtraction coefficient from the amplitude spectrum of the input voice may result in a negative value in calculation. Also in this case, since the amplitude spectrum itself cannot be a negative value originally, the portion is clipped and set to zero (0) or a relatively small positive constant. Therefore, a specific distortion is caused in the amplitude spectrum of the input speech whose noise has been suppressed obtained by the spectrum subtraction method.

[0009] The noise suppressor 203 shown in FIG. 2A supplies the speech recognizer 204 with a power spectrum of the input speech noise-suppressed by the spectrum subtraction method or a self-phase obtained by performing an inverse Fourier transform of the power spectrum. The relation number is output. As described above, since the power spectrum or the autocorrelation coefficient input to the speech recognition device 204 is distorted, the recognition rate of the speech recognition device 204 is reduced.

In order to suppress such a decrease in the recognition rate, the amplitude spectrum obtained by subtracting the amplitude spectrum of the estimated noise output from the spectrum subtraction method, the power spectrum obtained by subtracting the power spectrum of the estimated noise, or the power spectrum obtained by subtracting the power spectrum is used. It is conceivable to correct the autocorrelation coefficient obtained by performing the inverse Fourier transform to reduce the distortion peculiar to the above-described spectrum subtraction method. At this time, considering that different types of noise are used in different environments in which the speech recognition system is used, it is desirable to perform correction in accordance with noise generated in the environment in which the speech recognition system is used. However, when a system is designed to perform correction in accordance with noise generated in an assumed use environment, there is a concern that a recognition rate may be reduced when the system is used in an unexpected environment.

Therefore, the present invention suppresses noise by using a spectrum subtraction method for an input signal in which speech and noise to be recognized are mixed, regardless of the type of noise generated in any environment. It is an object of the present invention to optimally correct distortion generated at the time and reduce the distortion as much as possible, thereby contributing to improvement of a recognition rate in speech recognition.

[0012]

A speech recognition system according to the present invention, which has been made to achieve the above-mentioned object, comprises a noise suppressing apparatus for suppressing noise inputted to a voice input through a microphone or the like. And a speech recognition device connected to the noise suppression device and recognizing speech input from the noise suppression device.

[0013] In the noise suppression apparatus of the speech recognition system according to the present invention, the frame dividing means cuts out an input signal input through, for example, a microphone or the like as a frame signal at every predetermined processing time. The spectrum is calculated by performing a Fourier transform on the signal. Whether the input signal is a speech section in which speech is included or a noise section in which speech is not included is determined by the determination unit, and the noise spectrum estimation unit includes:
The noise spectrum is estimated using the spectrum calculated based on the input signal in the noise section. Then, the subtraction means subtracts a value obtained by multiplying the noise spectrum by a predetermined subtraction coefficient from the spectrum calculated based on the input signal of the voice section, and the clipping means subtracts a negative part of the subtraction result by the subtraction means to zero. Alternatively, a noise subtraction spectrum having a relatively small positive constant is calculated. Further, when the parameters of the neural network are properly adjusted, the mapping calculation means configured using the neural network model converts the noise subtraction spectrum calculated by the clipping means into distortion due to clipping. Map to reduced spectrum. Further, the noise subtraction spectrum calculated by the clipping means is stored and held in the noise subtraction spectrum storage means.

When the spectrum calculated by the mapping calculation means is input to the voice input device, the voice input device uses the recognition means to compare with the plurality of pre-stored comparison target pattern candidates with a high degree of matching. Is the recognition result. Further, the teacher spectrum storage means of the speech recognition device stores a noise-free teacher spectrum prepared for each comparison target pattern candidate,
The teacher spectrum output means of the speech recognition device outputs a teacher spectrum corresponding to the recognition result by the recognition means to the noise suppression device.

[0015] The learning control means of the noise suppression device, if the recognition result by the recognition means of the speech recognition device is correct, as described above, the teacher spectrum output from the speech recognition device corresponding to the output from the mapping calculation means. Is used as a teacher output of the neural network constituting the mapping calculation means, and a noise subtraction spectrum corresponding to the teacher spectrum stored and held in the noise subtraction spectrum storage means as described above is input to the neural network. Adjust network parameters.

The spectrum calculated by the spectrum calculating means may be an amplitude spectrum or a power spectrum. That is, when the frame signal is Fourier-transformed, the frequency spectrum S (f) is calculated. An amplitude spectrum A (f) which is an amplitude component of the frequency spectrum S (f) may be used, or a power spectrum P obtained by squaring the amplitude spectrum A (f).
(F) may be used.

For example, when the spectrum calculating means calculates the amplitude spectrum A (f), the noise spectrum estimating means estimates the noise amplitude spectrum AN (f), and the subtracting means based on the input signal of the voice section. From the calculated amplitude spectrum AS (f), a value obtained by multiplying the noise amplitude spectrum AN (f) by a predetermined subtraction coefficient is subtracted.

When the spectrum calculating means calculates the power spectrum P (f), the noise spectrum estimating means estimates the noise power spectrum PN (f), and the subtracting means calculates the power spectrum P (f) based on the input signal of the voice section. From the calculated power spectrum PS (f), a value obtained by multiplying the noise power spectrum PN (f) by a predetermined subtraction coefficient is subtracted.

When the subtraction process is performed as described above, a value obtained by multiplying the power spectrum or the amplitude spectrum of the estimated noise by the subtraction coefficient is subtracted from the power spectrum or the amplitude spectrum of the input voice, but when the subtraction coefficient is large. May be negative in calculations. Since the power spectrum or the amplitude spectrum cannot theoretically be a negative value, the portion is clipped to calculate a noise subtraction spectrum set to zero (0) or a relatively small positive constant. Therefore, a specific distortion due to clipping occurs in the noise subtraction spectrum.
If this is used as it is for speech recognition, the recognition rate will decrease.

Therefore, in the present invention, distortion reduction is realized by calculating a mapping from a spectrum in which distortion is caused by clipping to a spectrum in which the influence of clipping is corrected, using a neural network model. The neural network model has the ability to map any input signal (scalar, vector) to any output signal (scalar, vector). That is, an arbitrary input / output relationship can be realized by taking a sufficiently large number of neurons. Regarding this, for example, Kenichi Funabashi "Capability of neural network
About y ”(IEICE Technical Report vol.8
8, No. 126, MBE88-52). Therefore, if the parameters of the neural circuit model are adjusted, that is, if the neural circuit model is learned, the spectrum in which distortion has occurred due to clipping can be mapped to the spectrum in which the influence of clipping has been corrected.

As the neural network model, for example, a feedforward neural network model may be used. For example, input layer, hidden layer,
In the case of using a feedforward neural network model including three output layers, the input spectrum is sample data corresponding to a frequency f of 129 such as 1 to 129 with respect to a processing time t of a frame unit. If
It is conceivable to use one having 129 neurons in the input layer and the output layer. As a result, sample data corresponding to 129 frequencies f can be input from the input layer for each processing time t in frame units, and an output corresponding to 129 frequencies f can be obtained in the output layer. In this case, the noise subtraction spectrum input to the neural network is PSRC
(F, t) (f = 1, 2,..., 129) and the number of neurons in the hidden layer is, for example, 100, the output PSRCNN (f, t) (f = 1, 2, ...
129) is represented by the following equation.

[0022]

(Equation 1)

Here, W (f, h) (f = 1, 2,...)
, 129, h = 1, 2, ... 100), w (h,
i) (h = 1, 2,..., 100, i = 0, 1, 2,
, 129) are parameters of the neural network model. The function S (•) is S (x) = 1 / (1 + e ^−x ).

The parameters W (f, h) and w (h,
By adjusting i), mapping from a spectrum in which distortion is caused by clipping to a spectrum in which the influence of clipping is corrected is realized. By the way, considering that different types of noise are used in different environments where the speech recognition system is used, it is desirable to correct the influence of clipping in accordance with the noise generated in the environment where the speech recognition system is used. For example, when a speech recognition system using such a neural network model is used in a car, the parameters of the neural network are adjusted in advance by speech containing a lot of car noise. is there. However, if the system is designed to correct the distortion due to clipping in accordance with the noise generated in the assumed usage environment, the recognition rate will decrease due to the difference in the type of noise when used in an environment that was not assumed. It is concerned. For example, when a speech recognition system in which the parameters of a neural network are adjusted with speech including vehicle noise as described above is used in an information terminal device or the like set in a street or a parking area, for example, Depending on the type, the recognition rate may decrease.

On the other hand, in the speech recognition system of the present invention, a teacher spectrum that does not include noise is stored in the speech recognition device in correspondence with a plurality of comparison pattern candidates indicating the vocabulary to be recognized by the speech recognition device. Also, the noise subtraction spectrum input to the neural network is held. Then, when the speech recognition device performs correct recognition based on the noise subtraction spectrum, the teacher spectrum corresponding to the recognition result is used as the teacher output of the neural network, and the noise subtraction spectrum corresponding to the teacher spectrum is used as the neural circuit. As the input of the network, the parameters of the neural network are adjusted and the neural network is learned. In other words, if the parameters of the neural network were adjusted so that the spectrum distorted due to clipping was mapped to the teacher spectrum, which is the ideal output, the optimal correction would be made, so the correct recognition was made from the mapping of the distorted spectrum. In this case, the parameters are adjusted using the stored teacher spectrum as the teacher output.

As described above, in the speech recognition system of the present invention, the parameters of the neural network are adjusted each time correct speech recognition is performed during operation of the system. Therefore, even when used in any environment where the type of generated noise is different, the distortion generated when performing noise suppression using the spectrum subtraction method on the input signal in which speech and noise to be recognized are mixed is appropriate. Can be corrected.
In addition, since the neural network model can realize an arbitrary mapping, if parameter adjustment of the neural network, that is, learning progresses, a distorted spectrum can be mapped to a spectrum close to the teacher spectrum, and the spectrum can be mapped. Nearly optimal correction for distortion can be realized. As a result, it is possible to contribute to improvement of the recognition rate in voice recognition.

As described above, in the present invention,
The neural network model is used because the neural network model uses arbitrary input signals (scalar, vector).
Is mapped to an arbitrary output signal (scalar, vector), and has a well-known learning ability. That is, in the neural network model, the parameter adjustment, that is, learning can be performed based on the input signal and the teacher output signal corresponding to the input signal.
There is no need for a human to grasp the causal relationship existing between the input signal and the output signal. As a result, by using the neural network model, the distortion of the spectrum due to clipping can be reduced in accordance with any noise generated in various environments where the speech recognition system is used.

The power spectrum PS calculated by the subtracting means based on the input signal of the voice section.
If it is assumed that a configuration obtained by multiplying the noise power spectrum PN (f) by a predetermined subtraction coefficient from (f) is to be subtracted, a configuration as claimed in claim 4 may be adopted. That is, the noise suppression device is configured to include an autocorrelation coefficient calculation unit, and the autocorrelation coefficient calculation unit calculates an autocorrelation coefficient based on the noise subtraction spectrum calculated by the clipping unit. At this time, the mapping calculation means
The autocorrelation coefficient calculated by the autocorrelation coefficient calculation means is mapped to an autocorrelation coefficient with reduced distortion due to clipping.

As described above, the distortion can be similarly reduced by using the autocorrelation coefficient, and the memory capacity and the processing load in the speech recognition apparatus for performing the speech recognition using the output from the noise suppression apparatus can be reduced. It is valid. This focuses on the fact that the Fourier transform of the autocorrelation coefficient becomes a power spectrum, that is, the inverse Fourier transform of the power spectrum becomes an autocorrelation coefficient. Assuming that the autocorrelation coefficient is C (r, t) and the inverse Fourier transform is F- ¹ , the relationship with the power spectrum P (f, t) is as follows.

C (r, t) = F ^-1 [P (f, t)] Here, r is an index of the autocorrelation coefficient, and corresponds to the frequency f in the power spectrum. As described above, since the power spectrum and the autocorrelation coefficient are equivalent, if the mapping of the autocorrelation coefficient is calculated, the same result as when the power spectrum is used, that is, an output with reduced distortion is obtained. be able to.

The use of such an autocorrelation coefficient reduces the memory capacity and processing load of the speech recognition apparatus. This will be described. The speech recognizer uses linear predictive coding (LP)
C), and assuming that the power spectrum is output from the noise suppression device, the speech recognition device must first calculate the autocorrelation coefficient from the power spectrum output from the noise suppression device. . Therefore, the processing load and the memory capacity are increased. On the other hand, if the noise suppression device converts the autocorrelation coefficient to the speech recognition device and passes it to the speech recognition device, the processing load and the memory capacity of the speech recognition device can be reduced.
When the speech recognizer performs the P-order LPC, the autocorrelation coefficient C (r, r) of the index r is r = 0, 1, 2,.
t), and generally P = about 17.

Therefore, the power spectrum can be inversely Fourier-transformed into autocorrelation coefficients, and a map of the autocorrelation coefficients is output, thereby reducing the memory capacity and processing load in the speech recognition device. By the way, in the speech recognition system of the present invention, when the recognition result of the speech recognition device is correct, the parameters of the neural network are adjusted. Whether the recognition result is correct is determined, for example, as described in claim 6. It is conceivable to notify the recognition result by the recognition means and determine whether or not the recognition result is correct based on a confirmation signal input from outside in response to the notification. Here, the notification of the recognition result is performed by, for example, outputting the recognition result as a synthesized voice, or displaying the recognition result on a display of a system having a display such as a navigation system. Then, for example, the speaker itself determines whether or not the notified recognition result is correct, and presses a confirmation switch or performs a predetermined voice input. It is determined whether or not the recognition result is correct based on an input confirmation signal such as an audio signal.

Further, for example, when there is a difference of a predetermined value or more between the matching degree of a pattern determined to have a high matching degree by the recognizing means and the matching degree of another comparison target pattern candidate. Alternatively, a configuration may be adopted in which the recognition result by the recognition means is determined to be correct. The recognition means of the speech recognition device calculates the degree of coincidence with the output from the noise suppression device for each of a plurality of pre-stored comparison target pattern candidates, and uses the one with a high degree of coincidence as the recognition result. If there is no difference between the calculated degrees of coincidence, there is a high possibility that the degree of recognition will be erroneously recognized. Therefore, when the degree of coincidence of the pattern as the recognition result is relatively large from the degree of coincidence of the other patterns, it is estimated that the pattern is correctly recognized. This is convenient because the user does not need to confirm the correctness of the recognition result as described above.

The above-described determination means determines whether the input signal is a voice section in which speech is included or a noise section in which no speech is included. This is based on the power of the input signal. It is conceivable that the determination is made based on this. Also, the input period specified by the input period specifying means provided for the speaker to specify the period for inputting the voice may be determined as the voice section. For example, PTT (Push-To-Ta)
lk) switch. In other words, if the user
Pressing the T switch while inputting a voice, the PTT
The voice input while the switch is pressed is received as a processing target. By doing so, it is only necessary to execute the noise suppression processing only on the input signal to be subjected to noise suppression, which is effective in reducing the processing load.

Such a speech recognition system can adaptively correct a specific distortion by the spectrum subtraction method under any environment where the type of noise mixed into the speech signal is different. You can think. For example, it can be used for a so-called car navigation system. In this case,
For example, it is very convenient if a destination for route setting can be input by voice. It is also conceivable to use not only a navigation system but also a voice recognition system for an in-vehicle air conditioning system, for example. In this case, the air-conditioning system-related instructions in the air-conditioning system are used by the user to input by voice. Further, for example, the present invention can be similarly applied to a portable information terminal device or an information terminal device set in a street or a parking area.

[0036]

FIG. 1 is a block diagram showing a schematic configuration of a speech recognition system according to an embodiment of the present invention. The speech recognition system includes a noise suppression device 10 that performs noise suppression on speech input via a microphone 30, and outputs an output from the noise suppression device 10 to a plurality of comparison target pattern candidates stored in advance. And a speech recognition device 20 that recognizes a result having a higher degree of coincidence as a recognition result.

First, the noise suppression device 10 will be described. As shown in FIG. 1, the noise suppression device 10 includes a voice input unit 11, an input voice cutout unit 12, a frame power spectrum calculation unit 13, a noise power spectrum estimation unit 14, a noise power spectrum subtraction unit 15, It includes a power spectrum clipping calculator 16, a mapping calculator 17, and a learning controller 18. Hereinafter, processing contents in each block will be described.

The audio input section 11 converts an analog audio signal input via the microphone 30 into a digital signal at a sampling frequency of, for example, 10 KHz, and outputs the digital signal to the input audio cutout section 12. The input audio cutout unit 12 sequentially cuts out the input signal from the audio input unit 11 for each frame of a predetermined length and a predetermined sample, for example, 256 samples in 25.6 milliseconds.
3 and the noise power spectrum estimating unit 14.

The noise power spectrum estimating unit 14 Fourier-transforms the frame signal from the input speech extracting unit 12 to obtain a spectrum SN (f, t), and calculates the square of the amplitude of the spectrum SN (f, t). The calculated power spectrum PN (f, t) is stored in an internal buffer (not shown). This power spectrum PN
The formula for calculating (f, t) is as follows.

[0040]

(Equation 2)

This buffer does not store all the past power spectra, but only 5 buffers from the current time.
× 25.6 ms in the past power spectrum, that is, the power spectrum P for the latest five frames
N (f, t) is stored while being sequentially updated.

The power spectrum PN (f, t)
In the equation, f is a frequency, and t is a time (in this case, a time corresponding to processing in a frame unit). Since the frame signal of 256 samples is Fourier-transformed, the power spectrum PN (f, t) is 129 in the frequency (f) direction due to the symmetry of the spectrum calculated by the Fourier transform.
Sample data. Also, t = 0 is now t =
It is assumed that the more the number increases, the more the past, such as 1 indicates the immediately preceding past, t = 2 indicates the previous past, and so on. Therefore, the power spectra PN (f, t) for the latest five frames are PN (f, 0), PN (f, t).
1), PN (f, 2), PN (f, 3), PN (f,
4) (here, f = 1, 2,..., 129). The previous power spectrum before that is discarded from the buffer.

Then, the noise power spectrum estimating section 1
4 receives the voice input detection signal indicating that the voice has been input, and stops the process of estimating the noise power spectrum. In the present embodiment, a PTT (Push-T
When the o-Talk) switch is pressed, this voice input detection signal is output. That is, in the present voice recognition system, the user inputs a voice via the microphone 30 while pressing the PTT switch. Therefore, when the PTT switch is pressed, it means that the user has operated with an intention to input a voice. In this case, the user does not need to determine whether or not there is a voice input. Processing is perceived as a period during which an input is made (voice section).

The noise power spectrum estimating section 14 which has received the voice input detection signal stops the noise power spectrum estimating process, and the five power spectra PN (f, 0) and PN (f, 1) stored in the buffer. ),
The average value of PN (f, 2), PN (f, 3) and PN (f, 4) is calculated, and the noise power spectrum PNAV (f) used for subtraction in the spectrum subtraction method
(F is a frequency) and passes it to the noise power spectrum subtraction unit 15. Note that this noise power spectrum PN
The formula for calculating AV (f) is as follows.

[0045]

(Equation 3)

On the other hand, the frame power spectrum calculator 13 performs processing only when receiving a voice input detection signal. In the processing, a Fourier transform is performed on the frame signal from the input voice cut-out unit 12 and the input voice signal for each 25.6 millisecond frame as described above to obtain a spectrum SS (f, t) of the input voice signal. , And calculates the square of the amplitude of the spectrum SS (f, t) to obtain a power spectrum PS (f, t), which is passed to the noise power spectrum subtraction unit 15. Note that this power spectrum P
The formula for calculating S (f, t) is as follows.

[0047]

(Equation 4)

In the noise power spectrum subtractor 15,
The power spectrum PS (f, t) sent from the frame power spectrum calculator 13 is subtracted by multiplying the noise power spectrum PNAV (f) sent from the noise power spectrum estimator 14 by a predetermined subtraction coefficient. Power spectrum clipping calculator 1
Send to 6. Here, the subtract coefficient is 1.2. Therefore, the output P from the noise power spectrum subtractor 15 is
SR (f, t) is as shown in the following equation.

[0049]

(Equation 5)

Power spectrum clipping calculator 1
6, the power spectrum PSRC (f, t) obtained by clipping the portion that has become negative by the subtraction to the output PSR (f, t) from the noise power spectrum subtractor 15 and replacing it with zero or a relatively small positive constant.
t) is calculated and passed to the mapping calculation unit 17. Therefore, the power spectrum PSRC (f, t) is as shown in the following equation.

[0051]

(Equation 6)

The mapping calculation unit 17 is configured using a three-layer feed-forward type neural network model. The input layer, the hidden layer, and the output layer are composed of 129, 100, and 129 neurons, respectively. The mapping calculation unit 17 receives the power spectrum PSRC (f, t) from the power spectrum clipping calculation unit 16 and receives the power spectrum PRC (f, t).
SRCNN (f, t) is output. Power spectrum PSRC (f, t), which is time-series data in frame units
Is data of 129 samples in the frequency (f) direction, and is input to this neural network model,
PSRCNN (f, t), which is 129 samples of time-series data, is output every time t corresponding to processing in units of frames. The output PSRCNN (f,
t) is as shown in the following equation.

[0053]

(Equation 7)

Here, W (f, h) (f = 1, 2,...)
, 129, h = 1, 2, ... 100), w (h,
i) (h = 1, 2,..., 100, i = 0, 1, 2,
, 129) are parameters of the neural network model. The function S (•) is S (x) = 1 / (1 + e ^−x ).

The mapping calculation section 17 sequentially stores the power spectrum PSRC (f, t) output from the power spectrum clipping calculation section 16 in the buffer 17a while the PTT switch is pressed (voice section). In this way, the power spectrum PSRCNN (f) obtained for each frame of 25.6 milliseconds, which is a cutout unit in the input speech cutout unit 12, is sequentially sent to the speech recognition device 20.

Next, the speech recognition device 20 will be described. The speech recognition device 20 includes an inverse Fourier transform unit 21;
An LPC analysis unit 22, a cepstrum calculation unit 23, a standard pattern storage unit 24, a collation unit 25, and a determination unit 26 are provided.

In the inverse Fourier transform unit 21, the mapping calculation unit 1
7 is subjected to an inverse Fourier transform on the output PSRCNN (f, t) to obtain an autocorrelation coefficient.
Pass to In the LPC analysis unit 22, the inverse Fourier transform unit 21
A linear prediction analysis is performed using the output from. Linear prediction analysis is a common analysis technique in the field of speech signal processing. For example, Furui "Digital Speech Processing" (Tokai University Press)
And so on. In the present embodiment, the autocorrelation method is used for the linear prediction analysis, and a 12th-order LPC coefficient is calculated using the autocorrelation coefficient.

Then, the cepstrum calculator 23 calculates L
Based on the LPC coefficient calculated by the PC analysis unit 22,
The LPC cepstrum coefficient as a feature parameter on the spectrum for each frame is calculated from the first order to the 17th order. On the other hand, a standard pattern (feature parameter sequence) of the vocabulary to be recognized, which has been calculated in advance, is stored in the standard pattern storage unit 24, and the matching unit 25 compares the standard pattern stored in the standard pattern storage unit 24 with the standard pattern. , And the LPC cepstrum coefficient calculated by the cepstrum calculation unit 23. For these, a similarity corresponding to a word stored as dictionary data is obtained by a known DP matching method, an HMM (Hidden Markov Model), a neural network model, or the like. Then, the determination unit 26
The vocabulary having the highest similarity calculated by the matching unit 25 among the vocabularies to be recognized is output as a recognition result. Note that the speech recognition device 20 of the present embodiment performs recognition in units of words as described above, but may perform recognition in units of phrases or sentences.

As described above, according to the speech recognition system of the present embodiment, the noise power spectrum subtracting section 15 of the noise suppression device 10 uses the power spectrum PS (f, t) calculated based on the input signal of the speech section. The noise spectrum PNAV (f) multiplied by a predetermined subtraction coefficient (here, 1.2) is subtracted, and the mapping calculation unit 17 calculates PSRCNN (f, t).
The noise power spectrum subtracting section 15 subtracts 1.2 times the noise power spectrum PN (f) from the power spectrum PS (f, t) calculated based on the input signal in the voice section as described above. ing. In this case, since the subtract coefficient is 1.2, which is larger than 1, it may be a negative value in calculation. Since the power spectrum cannot theoretically be a negative value, the power spectrum clipping calculation unit 16 clips that part and sets it to zero (0) or a relatively small positive constant. Therefore, a distortion peculiar to the noise-suppressed power spectrum output from the power spectrum clipping calculation unit 16 occurs, and if this is used for speech recognition as it is, the recognition rate is reduced.

Therefore, distortion is reduced by calculating a mapping from a power spectrum in which distortion is caused by clipping to a power spectrum in which the influence of clipping is corrected, using a neural network model. The neural network model has the ability to map any input signal (scalar, vector) to any output signal (scalar, vector). That is, an arbitrary input / output relationship can be realized by taking a sufficiently large number of neurons. Regarding this, for example, Kenichi Funabashi, “Capabili of Neural Network
About ty "(IEICE Technical Report vol.8
8, No. 126, MBE88-52). Therefore, if the parameters of the neural network model are adjusted, that is, if the neural network model is appropriately learned, the mapping calculation unit 17 obtains a power spectrum in which distortion due to clipping has been reduced from the power spectrum in which distortion has occurred due to clipping. Can be calculated.

Next, adjustment of the parameters of the neural network model constituting the mapping calculation unit 17 of the noise suppression device 10 in the speech recognition system of the present embodiment will be described. As described above, in the speech recognition device 20, the determination unit 26
Outputs the vocabulary having the highest similarity calculated by the matching unit 25 among the vocabularies to be recognized as a recognition result.
The recognition result of the determination unit 26 is output to a speech synthesis unit 41 connected to the outside of the speech recognition device 20.

The voice synthesizing section 41 synthesizes the output from the judging section 26 as a digital signal having a predetermined frequency, and the voice output section 42 converts this digital signal into an analog voice signal and outputs it. As a result, the word recognized from the speaker 43 is output as a voice.

Thereafter, the determination section 26 waits for the input of a confirmation signal from the outside. In the present embodiment, a confirmation signal is output when a confirmation switch (not shown) is pressed. That is, in the present voice recognition system, the user presses the confirmation switch when the word output from the speaker 43 is correctly recognized. Accordingly, when the confirmation signal is input to the determination unit 26, it means that the voice input by the user has been correctly recognized, and in that case, the operation proceeds to a learning operation to be described later. On the other hand, when the confirmation signal is not input to the determination unit 26, it means that the voice input by the user is incorrectly recognized, and in that case, the learning operation is not performed.

Here, the learning operation in the speech recognition system of the present embodiment will be described. The determination unit 26 is provided with a buffer 26a, which stores a teacher power spectrum including no noise corresponding to each vocabulary to be recognized. Then, when the confirmation signal is input as described above, the determination unit 26 outputs the teacher power spectrum corresponding to the recognition result to the learning control unit 18 of the noise suppression device 10.

As described above, the power spectrum PSRC (f, t) of the period (voice section) during which the PTT switch is pressed is stored in the buffer 17a of the mapping calculation unit 17. Therefore, the power spectrum PSRC (f, t) stored in the buffer 17a or a part thereof corresponds to the recognized word.

Therefore, the learning control unit 18 uses the teacher power spectrum corresponding to the recognition result output from the determination unit 26 as the teacher output of the neural network of the mapping calculation unit 17, and uses the power spectrum PSRC (f, t) (t Is used as an input of a neural network of the mapping calculation unit 17 to adjust parameters of the neural network. This adjustment can be performed by, for example, a back propagation learning method (Technical Review, “Neurocomputer”, Chapter 2, Back Propagation pp. 28-pp.
84).

Here, a specific example of a learning procedure of the neural network constituting the mapping calculation unit 17 of the above-described noise suppression apparatus 10 of the present embodiment will be described. First, parameters are adjusted so that the neural network outputs the input signal as it is without changing the value of the input signal. About this, for example, 10 men,
A plurality of power spectra are created from a plurality of human voices such as 10 women, and parameters are adjusted by a back propagation learning method using each power spectrum as an input and a teacher output of a neural network.

The above-described mapping calculation unit 17 is constituted by the neural network model that has performed such learning. Accordingly, the noise suppression device 10 outputs the output PSRC (f, t) from the power spectrum clipping calculation unit 16 as it is. That is, the output PSRCNN (f, t) from the mapping calculation unit 17 is equal to the input PSRC (f, t) to the mapping calculation unit 17. That is, the power spectrum PSRCNN (f, t) in which the influence of clipping is not corrected at all is output to the speech recognition device 20.

However, even in this case, although the recognition rate is lowered by the distortion due to clipping, the recognition may be performed correctly. If recognized correctly,
As described above, the teacher output of the neural network is used as the teacher power spectrum stored in the buffer 26a of the determination unit 26 of the speech recognition device, and the input of the neural network is used as the mapping calculation unit 17.
Power spectrum PS stored in buffer 17a
The parameter is adjusted as RC (f, t). Then, each time the voice input from the user is correctly recognized, the adjustment of the repetition parameter is performed. If the parameters are adjusted, the recognition rate is improved, so that learning is performed more frequently.

As described above, in the speech recognition system of this embodiment, the parameters of the neural network of the mapping calculation unit 17 are adjusted each time correct speech recognition is performed during operation of the system. Therefore, even when used in any environment where the type of generated noise is different, the distortion generated when performing noise suppression using the spectrum subtraction method on the input signal in which speech and noise to be recognized are mixed is appropriate. Can be corrected. In addition, since the neural network model can realize an arbitrary mapping, if the parameter adjustment of the neural network constituting the mapping calculation unit 17, that is, learning progresses, the distorted spectrum is mapped to a spectrum close to the teacher spectrum. , And a near-optimal correction for spectrum distortion can be realized. As a result, it is possible to contribute to improvement of the recognition rate in voice recognition.

In the present embodiment, the cut-out function of the input sound cut-out unit 12 corresponds to “frame division means”. In addition, in the input speech clipping unit 12, when a speech input detection signal is input, a clipping process starts, or the noise power spectrum estimating unit 14
The estimation of the noise power spectrum is stopped when a voice input detection signal is input, but this corresponds to a change in the processing content based on the determination result of the voice section and the noise section by the "determining means". The frame power spectrum calculator 13 corresponds to “spectrum calculator”, and the noise power spectrum estimator 14 corresponds to “noise spectrum estimator”. Further, the noise power spectrum subtraction unit 15 corresponds to “subtraction means”, the power spectrum clipping calculation unit 16 corresponds to “clipping means”,
The mapping calculation unit 17 corresponds to “mapping calculation means”. The buffer 17a of the mapping calculation unit 17 corresponds to "noise subtraction spectrum storage means".

The determination unit 26 of the voice recognition device 20 corresponds to “recognition means” and “teacher spectrum output means”, and the buffer 26a of the determination unit 26 corresponds to “teacher spectrum storage means”. As described above, the present invention is not limited to such an embodiment at all, and can be implemented in various forms without departing from the gist of the present invention.

(1) For example, in the above embodiment,
Power spectrum P output from the mapping calculation unit 17
Although SRCNN (f, t) is passed to the inverse Fourier transform unit 21 of the speech recognition device 20, the noise suppressor 10 is provided with an inverse Fourier transform unit, and the power which is the output from the power spectrum clipping calculation unit 16 The spectrum PSRC (f, t) may be inverse Fourier-transformed into autocorrelation coefficients, and then the mapping calculation may be performed by the mapping calculation unit 17. As described above, the use of the autocorrelation coefficient can also reduce the distortion, and in this case, it is effective in reducing the memory capacity and the processing load in the subsequent speech recognition device 20.

This is the inverse Fourier of the power spectrum.
It focuses on the fact that the conversion becomes an autocorrelation coefficient.
That is, the autocorrelation coefficient is C (r, t), and the inverse Fourier transform
To F ^-1Then, the power spectrum P (f, t)
The relationship is as follows: C (r, t) = F^-1[P (f, t)] Here, r is an index of the autocorrelation coefficient, and the power spectrum
It corresponds to the frequency f in the ram.

As described above, since the power spectrum and the autocorrelation coefficient are equivalent, the same result can be obtained by using the power spectrum and the autocorrelation coefficient, that is, an output with reduced distortion can be obtained. it can. If the noise suppression device 10 performs autocorrelation coefficient conversion as described above, the speech recognition device 20 does not need to perform the processing by the inverse Fourier transform unit 21.
Thus, the processing load and the memory capacity can be reduced.

(2) In the above embodiment, the power spectrum PS (f) obtained by squaring the amplitude of the frequency spectrum S (f) obtained by Fourier transform is used, and the noise power spectrum PN (f) is similarly obtained. However, the amplitude spectrum A (f) itself, which is the amplitude component of the frequency spectrum S (f), may be used. In that case, the noise amplitude spectrum AN (f) is estimated,
From the amplitude spectrum AS (f) calculated based on the input signal of the voice section, a noise amplitude spectrum ANAV (f)
Is multiplied by a predetermined subtraction coefficient.

However, since the autocorrelation coefficient C (r, t) is equivalent to the power spectrum P (f, t) as described above, when the amplitude spectrum is used, the autocorrelation coefficient C (r, t) Therefore, the advantage of using the autocorrelation coefficient cannot be obtained. However, when considered conversely, the reason why the noise suppression apparatus 10 converts the autocorrelation coefficient into an autocorrelation coefficient is that when the autocorrelation coefficient is passed to the speech recognition apparatus 20, the processing load and the memory capacity of the speech recognition apparatus 20 can be reduced. If this advantage is not enjoyed, it is not necessary to convert to an autocorrelation coefficient.

(3) In the above embodiment, the mapping calculation unit 17 calculates the mapping of the power spectrum PSRC (f) for each time t corresponding to the processing in units of frames. The mapping of the power spectrum PSRC (t) for a predetermined period may be calculated. In addition,
As shown in (1), when the autocorrelation coefficient is used,
The mapping of the autocorrelation coefficient CSS (r) may be calculated for each time t corresponding to the processing in units of frames, or the mapping of the autocorrelation coefficient CSS (t) for a predetermined period may be calculated for each index r. You may.

(4) Furthermore, in the above embodiment, the correctness / incorrectness of the recognition result by the determination unit 26 of the voice recognition device 20 is performed based on a confirmation signal input from the outside. That is, the recognition result is output from the speaker 43 as a synthesized voice, and the speaker confirms itself. If the recognition result is correct, the determination is made based on the confirmation signal input by pressing the confirmation switch by the speaker.

On the other hand, for example, the judgment unit 26 compares the similarity calculated by the matching unit 25 with each vocabulary to be recognized, and when the difference between the highest similarity and the next highest similarity is equal to or larger than a predetermined value. In some cases, the recognition may be determined to be correct. For example, if there are three vocabularies to be recognized, A, B, and C, and the similarities of vocabularies A, B, and C are 90%, 20%, and 30%, the recognition result of vocabulary A may be correct. This is because although the vocabulary A, B, and C have similarities of 50%, 45%, and 40%, the possibility that the recognition result of the vocabulary A is correct is low. As described above, if the determination unit 26 performs the correctness determination of the recognition result, it is convenient that the speaker does not need to perform the correctness determination of the recognition result as described above.

(5) In the above embodiment, when a user inputs a voice while pressing the PTT switch using a PTT switch provided for the speaker himself to designate a period for inputting the voice, While a period during which the PTT switch is pressed is regarded as a voice section, a voice section and a noise section may be determined based on an actual input signal. For example, it is conceivable to make the determination based on the power of the input signal.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a schematic configuration of a speech recognition system according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing an outline of a conventional speech recognition system.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 10 ... Noise suppression apparatus 11 ... Audio input part 12 ... Input audio cutout part 13 ... Frame power spectrum calculation part 14 ... Noise power spectrum estimation part 15 ... Noise power spectrum subtraction part 16 ... Power spectrum clipping calculation part 17 ... Mapping calculation part 17a ... Buffer 18 ... Learning control unit 20 ... Speech recognition device 21 ... Inverse Fourier transform unit 22 ... LPC
Analysis unit 23 ... Cepstrum calculation unit 24 ... Standard pattern storage unit 25 ... Collation unit 26 ... Decision unit 26a ... Buffer 30 ... Microphone 41 ... Speech synthesis unit 42 ... Speech output unit 43 ... Speaker 200 ... Speech recognition system 201 ... Microphone 203 ... Noise suppression device 204 ... Speech recognition device 205 ... PTT switch

Claims (9)

[Claims] 1. A frame dividing means for cutting out an input signal as a frame signal at every predetermined processing time, a spectrum calculating means for calculating a spectrum from the frame signal, a voice section in which voice is included in the input signal, and Determining means for determining a noise section that does not include voice, and a noise spectrum estimating means for estimating a noise spectrum using the spectrum calculated based on the input signal of the noise section determined by the determining means,
Subtraction means for subtracting a value obtained by multiplying the noise spectrum estimated by the noise spectrum estimation means by a predetermined subtraction coefficient from the spectrum calculated based on the input signal of the voice section; and A noise suppression device having a clipping unit that calculates a noise subtraction spectrum that is a zero or a relatively small positive constant by clipping a negative portion, and an output from the noise suppression device is stored in advance. A speech recognition system having a recognition unit that recognizes a pattern with a high degree of coincidence with a plurality of comparison target pattern candidates as a recognition result, wherein the speech recognition device further comprises: each of the comparison target pattern candidates Teacher spectrum that stores the teacher spectrum without noise prepared for A tram storage unit, and a teacher spectrum output unit that outputs a teacher spectrum corresponding to the recognition result by the recognition unit to the noise suppression device. On the other hand, the noise suppression device further includes: A noise subtraction spectrum storage means for storing a noise subtraction spectrum, and a neural network model capable of mapping the noise subtraction spectrum calculated by the clipping means to a spectrum in which distortion due to the clipping is reduced, A mapping calculating means for calculating a mapping of the noise subtraction spectrum; and, if the recognition result by the recognition means of the speech recognition device is correct, a teacher spectrum output from the speech recognition device corresponding to the output from the mapping calculation device. Neural network constituting the mapping calculation means Learning control means for adjusting the parameters of the neural network, using the noise subtracted spectrum corresponding to the teacher spectrum as an input to the neural network, as the teacher output of the neural network. A speech recognition system characterized by the following. 2. The speech recognition system according to claim 1, wherein the spectrum is an amplitude spectrum that is an amplitude component of a frequency spectrum obtained by performing a Fourier transform on a frame signal. 3. The speech recognition system according to claim 1, wherein the spectrum is a power spectrum obtained by squaring an amplitude spectrum which is an amplitude component of a frequency spectrum obtained by performing a Fourier transform on a frame signal. A speech recognition system characterized by the following. 4. The speech recognition system according to claim 3, further comprising: an autocorrelation coefficient calculation unit that calculates an autocorrelation coefficient based on the noise subtraction spectrum calculated by the clipping unit. The means is configured using a neural network model capable of mapping the autocorrelation coefficient calculated by the autocorrelation coefficient calculation means to an autocorrelation coefficient with reduced distortion due to the clipping. A speech recognition system characterized by the following. 5. The speech recognition system according to claim 1, wherein said mapping calculation means is configured using a feedforward neural network model. 6. The speech recognition system according to claim 1, wherein the recognition result is notified by the recognition unit, and the recognition result is determined based on a confirmation signal input from outside in response to the notification. A speech recognition system configured to determine whether or not is correct. 7. The speech recognition system according to claim 1, wherein the degree of coincidence of the pattern determined to be high in coincidence by the recognition unit and the degree of coincidence of another comparison target pattern candidate are determined. If there is a difference equal to or more than a predetermined value, the speech recognition system is configured to determine that the recognition result by the recognition unit is correct. 8. The speech recognition system according to claim 1, wherein the determination unit is configured to determine the voice section and the noise section based on the power of the input signal. Characteristic speech recognition system. 9. The speech recognition system according to claim 1, wherein the noise suppression device further comprises: an input period specifying unit provided for the speaker to specify a period during which the voice is input. A voice recognition system, wherein the determination unit is configured to determine the input period specified by the input period specification unit as the voice section.