JPH075895A - Device for recognition and method for recognizing voice in noisy evironment - Google Patents

Abstract

PURPOSE:To provide a voice recognition device in which noise components having rapid chagnes are exactly eliminated and a voice recognition rate is improved. CONSTITUTION:The device is provided with a microphone 11 which generates main signals ma in that voice signals sa and noise components oa . g from an audio device 16 are mixed, an amplifier 18 which generates reference signals ra based on the noise components, a voice section discrimination means which discriminates that voice section in which voice signal is contained in, the main signals or non-voice section containing no voice signal, a compensation coefficient updating means which generates and updates compensation coefficients based on the main signals ma in the non-voice section, a CPU 15 which has a computing means that subtracts the value obtained by multiplying the reference signals ra by the compensation coefficients in the voice section from the main signals ma and a voice recognition part 21 which collates the computation results obtained from the computing means and comparison voice signals B registered in a registration dictionary 22 and performs voice recognition.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a speech recognition apparatus for recognizing speech in a noisy environment and a speech recognition method in a noisy environment.

[0002]

2. Description of the Related Art In a conventional voice recognition device, in order to prevent a reduction in the recognition rate of voice recognition in such a noisy environment, an LMS method or a spectrum subtraction method (hereinafter referred to as "SS method") is used. And so on. The LMS method is a method of removing a noise component from a main signal, which is a microphone input signal in which vocalized voice and a noise component are mixed, by using a known noise signal as a reference signal by an adaptive filter method. Also, S. The S method is a method of removing noise components included in uttered speech by considering them as stationary noise.

FIG. 21 shows the S. It is a block diagram of the conventional speech recognition apparatus to which the S method is applied. In this figure, 1 receives a voice from a speaker (not shown) and a sound due to an audio signal which is a music noise from another source such as an audio device, and sends it as a main signal ma of an electric signal. It's Mike. Reference numeral 2 is an amplifier for amplifying the main signal ma. Reference numeral 3 is a filter bank for frequency-dividing the amplified main signal ma to generate a plurality of channel signals and selectively transmitting one of the channel signals.

The filter bank 3 is an all-pass filter 3a which passes the entire band of the main signal ma as a channel signal mo, and a plurality (n) of channel signals m1, m2, ..., Which divide the main signal ma into predetermined bands. mn
, A multiplexer for selecting one channel signal m (CH) (CH = 0,1,2, ..., n) of the bandpass filter groups 3b, 3c for transmitting the signal, and the channel signals ma, m1, m2 ,. 3d, the selected channel signal m (CH) is converted into a digital signal to obtain a channel signal M
It is composed of an A / D converter 3e that sends out (CH).

Reference numeral 4 denotes a CPU for recognizing the channel signal M (CH), which is the main audio data sent from the filter bank 3, by a voice, which is not shown in the drawing, but is not shown in the figure, but is a calculation unit, a ROM for storing a program, a data storing RAM and the like. Reference numeral 5 is a registration dictionary that stores comparative voice data registered in advance and supplies the comparative voice data to the CPU 4 during voice analysis.

Next, the operation of the conventional speech recognition apparatus will be described. Main signal m including audio noise component
After passing through the filter bank 3, a is converted into a digital signal for each channel and taken into the CPU 4 as main voice data to be voice-recognized. After that, the audio noise component is removed as a known noise component, and the voice recognition is performed by pattern matching with the comparative voice data registered in advance in the registration dictionary.

[0007]

However, in the above-mentioned conventional speech recognition apparatus, in the case of the LMS method, that is, the adaptive filter method, it responds to a too rapid change with respect to non-stationary noise such as an audio noise component. Even if it is not possible and there is no sudden change in noise, there is a problem that the convergence time of the filter becomes long. Further, since a processing device capable of high-speed arithmetic processing such as a DSP is required, there is a problem that it causes a cost increase of the voice recognition device.

In addition, S. Also in the case of the S method, there is a problem that the recognition rate cannot be increased because noise cannot be accurately removed from noise accompanied by a sudden change such as an audio noise component.

The speech recognition apparatus according to the present invention solves such a conventional problem, and accurately corrects a noise component accompanied by an abrupt change without requiring an expensive high-speed arithmetic processing device such as a DSP. It is an object of the present invention to perform excellent speech recognition by removing the speech and improving the speech recognition rate.

Another object of the speech recognition method according to the present invention is to further improve the effect of the adaptive SS system by using fuzzy inference.

[0011]

In order to achieve the above-mentioned object, a speech recognition apparatus according to the present invention is a comparison in which the noise component is removed from a main signal in which a voice signal from a speaker is mixed with a noise component and is registered in advance. A voice recognition device for recognizing the voice signal by collating with the voice signal, comprising means for generating a reference signal based on the noise component, and a voice section in which the main signal includes the voice signal or not. A voice section discriminating means for discriminating whether or not it is a non-voice section, a correction coefficient updating means for generating and updating a correction coefficient based on the main signal in the non-voice section, and a correction coefficient for the reference signal in the voice section. An arithmetic means for subtracting the multiplied value from the main signal, and a recognition means for collating the arithmetic result obtained from the arithmetic means with the comparative voice signal for voice recognition are provided. And wherein the door.

In order to achieve the above object, the speech recognition method according to the present invention recognizes the voice of the speaker by removing the noise component from the input signal in which the noise component is mixed with the voice signal component from the speaker. A voice recognition method in a noise environment, wherein a voice section is detected from the input signal by fuzzy inference, and it is determined whether or not the noise component is mixed in this voice section, and according to the determination result. The correction count for predicting the audio signal component is updated, the updated correction count is adjusted, subtraction processing is performed based on the adjusted correction count, and voice recognition is performed using the subtraction result as the audio signal component. It is characterized by performing.

Further, it is a voice recognition method in a noise environment for recognizing the voice of the speaker by removing the noise component from an input signal in which an acoustic noise component and a running noise component are mixed in the voice signal component from the speaker. A correction count for detecting a voice section from the input signal, determining by fuzzy inference whether or not the running noise component is mixed in the voice section, and predicting the voice signal component according to the determination result. In the noise environment characterized by performing the adjustment of the updated correction count, performing subtraction processing based on the adjusted correction count, and performing voice recognition using the subtraction result as the voice signal component. Voice recognition method.

Furthermore, a voice recognition method for recognizing the voice of the speaker by removing the noise component from an input signal in which a noise component is mixed with a voice signal component from the speaker,
A voice section is detected from the input signal, it is determined whether or not the noise component is mixed in the voice section, and a correction count for predicting the voice signal component is updated according to the determination result, and fuzzy The updated correction count is adjusted by inference, subtraction processing is performed based on the adjusted correction count, and voice recognition is performed using the subtraction result as the voice signal component.

[0015]

Therefore, in the voice recognition apparatus according to the present invention, in the calculation for subtracting the value obtained by multiplying the reference signal by the correction coefficient from the main signal in the voice section, the calculation is performed while updating the correction coefficient. It is possible to accurately remove the component and improve the voice recognition rate.

Further, in the voice recognition method according to the present invention, the method of determining the voice trigger level, the method of determining the running noise determination level, and the method of determining the adjustment amount of the correction coefficient are performed by fuzzy inference, so that the adaptive S -The effect of the S method can be improved.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the first to seventh inventions will be described below in detail with reference to the drawings.

1. An embodiment of the first invention will be described.

FIG. 1 is a block diagram of a voice recognition apparatus according to the first embodiment of the present invention. In FIG. 1, 11 is a main signal ma which is an electric signal in response to a voice from a speaker (not shown) and a sound from an audio signal (which will be described later) from another source such as an audio device. It is a microphone to send as. Reference numeral 12 is an amplifier for amplifying the main signal ma. 13 is the amplified main signal m
a is frequency-divided to generate a plurality of channel signals,
It is a filter bank that selectively outputs the one channel signal.

The filter bank 13 is an all-pass filter 13a which passes the entire band of the main signal ma as a channel signal mo, and a plurality (n) of channel signals m1, m2, ..., Which divide the main signal ma into predetermined bands.
bandpass filter groups 13b and 13c for transmitting mn,
One of the channel signals ma, m1, m2, ..., Mn is a channel signal m (CH) (CH = 0, 1, 2, ..., N).
A multiplexer 13d for selecting the A / D converter 13e for converting the selected channel signal m (CH) into a digital signal and transmitting the channel signal M (CH).
Composed of.

Reference numeral 14 is a latch circuit for holding a channel signal M (CH) which is the main audio data sent from the filter bank 13. 15 is CP as a control means
U, which is a computing unit 15a serving as a computing unit, a correction coefficient storage unit 15b, and others, which are not shown in the figure, include a ROM for storing a program, a RAM for storing data, and the like.

Reference numeral 16 is an audio device, which sends out an audio signal oa. Reference numeral 17 is a speaker for converting the audio signal oa of the electric signal into sound. However, the audio signal transmitted from the speaker 17 becomes music noise that should be removed from the voice from the speaker.

Reference numeral 18 denotes an amplifier which amplifies the audio signal oa from the audio device 16 and outputs it as a reference signal ra which is a reference signal. Reference numeral 19 is a filter bank for frequency-dividing this reference signal ra to generate a plurality of channel signals and selectively transmitting one of the channel signals.

The filter bank 19 divides the reference signal ra into predetermined bands so as to pass the entire band of the reference signal ra as the channel signal ro, and divides the reference signal ra into a plurality (n) of channel signals r1.
, R2, ..., rn for transmitting bandpass filter group 1
9b, 19c, channel signals ra, r1, r2, ...
One of the channel signals r (CH) (CH =
0, 1, 2, ..., N) is selected by a multiplexer 19d, and the selected channel signal r (CH) is converted into a digital signal and an A / D converter 19e for transmitting a channel signal R (CH) is provided. It

Reference numeral 20 is a latch circuit for holding the channel signal R (CH) sent from the filter bank 19. Reference numeral 21 is a voice recognition unit as a recognition means for recognizing the voice analysis data of the channel signal S (CH) calculated and output by the CPU 15, and outputs the channel selection signal CS for selecting the channel signal to the CPU 15 and the multiplexer 13d. Supply to 19d. Reference numeral 22 is a registration dictionary that stores comparative voice data registered in advance and supplies the comparative voice data to the voice recognition unit 21 during voice analysis.

Reference numeral 23 is a voice trigger circuit as a voice section discriminating means for detecting the start of the voice section of the voice signal from the generator included in the main signal ma to generate a start signal (trigger signal) TR and supplying it to the CPU 15. Is.

Next, the correction coefficient stored in the correction coefficient storage section 15b will be described.

The main signal ma input from the microphone 11
Is represented by the following (Equation 1).

[0029]

[Equation 1] Here, sa receives the voice from the generator, and the microphone 11
Is a voice signal output as an electric signal by the microphone 1
The conversion characteristic of 1 is added. Further, oa is an audio signal transmitted from the audio device 16.
Further, g is a transmission characteristic in which the audio signal oa is converted into sound by the conversion characteristic of the speaker 17, and the generated sound is propagated and reaches the microphone 11.

The audio signal oa is sent to the audio device 16
Since it can be obtained more directly, if the transmission characteristic g is known,
The audio signal sa can be obtained from the following (Equation 2).

[0031]

[Equation 2] However, in order to obtain this transmission characteristic g, highly accurate measurement is required, and it is very difficult to obtain its accurate value.

Therefore, in this embodiment, a method of frequency-analyzing the main signal ma and the audio signal oa and obtaining the audio signal sa by using the digitized data is adopted.

The A / D converters 13e and 19e shown in FIG.
Digital signals M (CH) and R (CH) sent from
The following relationship (Equation 3) is established between the two.

[0034]

[Equation 3] However, this (Equation 3) is not always completely equal to the left term and the right term due to an error caused by digitizing the analog signal. In this (Equation 3), S (C
H) is data obtained by digitizing the audio signal sa, and G
(CH) is a correction coefficient for multiplying R (CH) to predict the audio signal component S (CH) included in the main signal M (CH).

From this (Equation 3), the audio signal S (CH)
Is represented by the following (Equation 4).

[0036]

[Equation 4] From this (Equation 4), the audio signal S (CH) included in the main signal M (CH) can be predicted.

This correction coefficient G (CH) can be estimated from the ratio of M (CH) and R (CH) when no sound is generated, if the number of channels n which is the frequency resolution is large. is there. That is, if S (CH) = 0,
This is because M (CH) = R (CH) · G (CH), and the correction coefficient G (CH) can be expressed as M (CH) / R (CH). The correction coefficient G (CH) calculated in the non-voice section is stored in the correction coefficient storage unit 15b.

FIG. 2 is a CPU 1 of the voice recognition apparatus shown in FIG.
6 is a flowchart showing the operation of FIG. This operation will be described below. The correction coefficient in this case is M in the non-voice section.
The cumulative values of (CH) and R (CH) for several seconds (here, one second) are set as ΣM (CH) and ΣR (CH), and the ratio thereof is used as a correction coefficient according to the following (Equation 5). To do.

[0039]

[Equation 5] In FIG. 2, the channel selection signal CS is fetched from the voice recognition unit 21 and output to the latch circuits 14 and 19 (step S11), and the latch timing and the latch circuit 14 are output.
And the timing to fetch data from 19 is created. After that, the data M (CH) and R (CH) are fetched from the latch circuits 14 and 19, and the voice trigger circuit 23
The trigger signal TR from is taken in (step S12).
The time point at which this trigger signal TR, that is, the start end signal is received is regarded as the voice start end, and the timer is set from that time point to 1.6.
A second (this section is the maximum allowable speech section length of the speech recognition unit 21) is a speech section.

Each time the data M (CH) and R (CH) is fetched, it is judged whether or not it is in the voice section (step S13).
If a section that is not a voice section continues for 1 second or longer, stock data M (CH) and R (C
H) is updated, and the cumulative values ΣM (CH) and ΣR are updated.
(CH) is calculated, the latest correction coefficient is created by equation (3), and the value of the correction coefficient in the correction coefficient storage unit 15b is updated (step S14).

On the other hand, in the case of the voice section, the (latest) correction coefficient G stored in the correction coefficient storage unit 15b.
The data of (CH) is read and the captured data R (C
H) to multiply the audio noise component R (CH)
G (CH) is calculated, and the main signal M (C
The audio noise component is subtracted from the data of (H) (step S15). The subtraction data that is the result of this subtraction is output as the data of the audio signal S (CH) (step S
16).

As described above, according to the first embodiment of the present invention, the latest correction coefficient is always obtained from the main signal and the reference signal in the non-voice section, so that the non-steady state such as the rapidly changing audio noise is generated. It is also possible to deal with noise, and the convergence time of the filter can be shortened when there is no sudden change in audio noise. Further, there is no need for an expensive processing device capable of high-speed arithmetic processing such as DSP.

Furthermore, since the reference signal does not include a voice signal, the estimation error can be reduced and high voice recognition can be performed even in an audio noise environment.

2. An embodiment of the second invention will be described.

A feature of the present invention is that when the correction coefficient is created and updated in the embodiment of the first invention, it is created and updated by using the delay data using the past data.

FIG. 3 is a block diagram of a voice recognition apparatus according to the first embodiment of the present invention. In FIG. 3, the first shown in FIG.
The same configurations as those of the embodiment of the invention are represented by the same reference numerals, and the description thereof will be omitted. As shown in FIG. 3, a voice trigger circuit is not provided in the configuration of this embodiment.
However, the CPU 15 is provided with a voice detecting section 15c which is a voice section determining means for detecting a voice section together with a calculating means for calculating data and a delay process.

In the voice detection section 15c, after the provisional start end of the voice signal is determined, a point in time past a certain time from the provisional start end is regarded as a definite start end and the voice section is determined. Therefore, the data delay process described below is performed.

FIG. 4 is a diagram showing how the voice start end of the voice recognition device shown in FIG. 3 is detected, and is a diagram showing how data is delayed. In FIGS. 4A and 4B, DP indicates the current value data of the voice waveform after removing the audio noise component. Although it is actually a digital signal, it is represented here as an analog signal for convenience of description. t
s is a tentative detection position of the voice start end, and a point at which a predetermined threshold level, which will be described later, is reached is defined as a detection position.

DD is voice waveform data obtained by delaying the present value data DP by a fixed time, that is, past value data. In the case of this embodiment, the delay time Td between the current value data DP and the past value data DD is 1 second. Therefore, the past value data DD is obtained by storing the present value data DP in the RAM (not shown) and then reading it one second later. The delay time Td is set to 1 second because it can be estimated that the maximum error time between the true start end and the provisional start end of the audio signal is 1 second.

T0 is the true voice section of the voice signal, which is 1.6 seconds in this case. T1 is a voice section judged by the CPU 15, and is 2.6 seconds in this case. Therefore,
In the case of FIG. 4A, the provisional start end ts of the audio signal is
This is a case where it is delayed by 1 second (maximum error time) from the true start point. Further, in the case of FIG. 4B, the provisional start end ts of the audio signal is delayed from the true start end by a slight time. In either case, the true voice section is included in the voice section determined by the CPU 15.

FIG. 5 is a flow chart showing the operation of the CPU 15 in the voice recognition device in the first embodiment of the second invention. The operation of the first embodiment of the present invention will be described below with reference to FIG.

First, the channel selection signal CS supplied from the voice recognition section 21 is monitored, and the latch circuits 14 and 2 are monitored.
The latch timing signal is output to 0 (step S2
1) The data M (CH) and R (CH) are fetched from the latch circuits 14 and 20 (step S22).

Next, the voice section is detected from the acquired data (step S23). Start t of this voice section
The detection of s is performed as follows. Filter bank 13
Using the correction coefficient before updating (which is referred to as G0 (0)) obtained from the all-pass filter 13a, the level of the audio signal becomes higher than the predetermined threshold level TH by the following (Equation 6). Point (time), that is, starting point t
s is detected.

[0054]

[Equation 6] In this (Equation 6), M (0) and R (0) are the main signal and the reference signal obtained from the all-pass filter 13a. As described above, the voice section T1 is set to 2.6 seconds from the start end ts.

Further, the stock data for the past 1 second from the starting end ts is updated. That is, the delay data MD (CH) and RD (CH) one second before are taken out from the RAM (step S24). By obtaining this delay data, the true voice section T0 of 1.6 seconds can be included in the voice section T1. Also, the delay data of the voice data is used because the correction coefficient is updated a few millimeters earlier than the voice start end of the delay data to reduce the residual component of the audio signal due to the estimation error before the voice start end,
To prevent the trigger at the beginning of the sound from being triggered too early 2
There are also secondary effects.

Next, it is judged whether or not it is in the voice section (step S25). If it is not in the voice section, the delay data MD
The correction coefficient calculation stock data is updated using (CH) and RD (CH) as the correction coefficient calculation data (step S26). Updated delay data MD (CH), RD
ΣMD (CH), ΣRD (C
H) is a cumulative value of delay data that does not include a voice component in the past 1 second. The correction coefficient G (CH) is updated by substituting this cumulative value into the following (Equation 7).

[0057]

[Equation 7] On the other hand, in step S25, in the case of the voice section, the updated correction coefficient G (CH) and the delay data MD
Using (CH) and RD (CH), the audio noise component is subtracted by the following (Equation 8) (step S27),
The subtracted data, that is, the delayed voice signal SD (CH) is obtained.

[0058]

[Equation 8] The delayed voice signal SD (CH) which is the subtracted data is output to the voice recognition unit 21 (step S28).

As described above, the present value data of the main signal and the reference signal, the voice section detection using the correction coefficient before the update, and the past value data of the main signal and the reference signal are used. The effect shown can be obtained.

1) Even if the audio noise is large, the audio noise component is removed in advance to some extent by the subtraction data using the pre-update correction coefficient, so that the detection error at the start end of the audio signal can be reduced.

2) Since the voice section is automatically detected, it is possible to eliminate the burden of the user performing an operation such as key input each time the user speaks.

3) Even if there is a residual component of audio noise component due to an estimation error because the level of the voice signal is small, the threshold level for detecting the voice signal is set to a large value, thereby erroneously detecting the voice section due to the audio noise. Can be reduced and an appropriate value of the correction coefficient can be obtained. Therefore, it does not depend on the threshold level of the voice recognition unit.

4) Real-time processing is possible because of the extremely simple method.

The block diagram of the speech recognition apparatus according to the second embodiment of the present invention has the same configuration as that of the second embodiment of the present invention shown in FIG. 3, and the description thereof will be omitted.

The feature of this embodiment is that the adaptive S.M. Instead of updating the correction coefficient in the S method every time when uttering,
It is to do it at regular intervals. When the voice recognition unit has a certain level of stationary noise removal function, if the correction coefficient is updated every time, the subtraction amount will change every time especially when the audio component greatly changes. As a result, the residual component of audio noise due to the estimation error changes every time,
The stationary noise removal function of the voice recognition part does not work effectively. Figure 6
(A) is a figure which shows the residual component of audio noise when a correction coefficient is updated every time it utters. In such a case, a phenomenon occurs in which the voice recognition unit is likely to erroneously detect a voice section. That is, this second embodiment is made to avoid the above-mentioned erroneous detection.

FIG. 7 is a flow chart showing the operation of the CPU 15 of the voice recognition device according to the second embodiment of the second invention. The operation of the third embodiment will be described based on this flowchart and FIG.

First, the channel selection signal CS from the voice recognition section 21 is monitored, the latch timing signal is supplied to the latch circuits 14 and 20 (step S31), and the fetch timing from the latch circuits 14 and 20 to the CPU 15 is determined. Structure, data M (CH), R (CH)
Is taken in (step S32).

The voice section is detected from the fetched data (step S33). This detection is performed by the correction coefficient G0 (0) before update obtained from the all-pass filter 13a.
Is used as the start point of the voice by the data satisfying (Equation 6).

[0069]

## EQU00006 ## The voice section is set to 2.6 seconds (maximum voice section length of the voice recognition device) from this voice start end.

Next, data stock and delayed data are taken out (step S34). Update the stock data for the past few seconds (1 second in this case) from the present,
As a result, delay data MD (CH) and RD (CH) are obtained. Then, it is determined whether or not it is a voice section (step S3
5) If it is not in the voice section, this delay data is stocked to create correction coefficient candidates (step S3).
6). Specifically, when it is not in the voice section, the obtained delay data MD (CH), RD (CH) is used as the correction coefficient calculation data, and the correction coefficient calculation stock data is updated. Then, the correction coefficient candidate Gc (CH) is obtained by the following (Equation 9) using the cumulative values ΣMD (CH) and ΣRD (CH) for the past one second that does not include the voice component.

[0071]

[Equation 9] After obtaining the correction coefficient candidate Gc (CH), the correction coefficient is updated at regular intervals (step S37). That is,
A counter is set, and this counter is incremented when it is not in a voice section, and a fixed time (in this case, 0.5
Every second), the correction coefficient candidate Gc (CH) is replaced with the correction coefficient G (C
H).

On the other hand, if it is not in the voice section, S36,
Omit S37. Next, the subtraction process of the audio noise component is performed regardless of the presence or absence of the voice section (step S38).
In this subtraction process, the audio signal SD (CH) is extracted by (Equation 8) using the updated correction coefficient G (CH) and the delay data MD (CH) and RD (CH), and the subtracted data is extracted. Output (step S39).

[0073]

## EQU00008 ## FIG. 6B is a diagram showing the residual component of the audio noise when the correction coefficient is updated at fixed time intervals (0.5 seconds). As is clear from this figure, since the fluctuation of the residual component is reduced, the residual component can be removed by the stationary noise removal function of the voice recognition unit.

Next, a third embodiment of the present invention will be described.

The block diagram of the speech recognition apparatus of this embodiment also has the same configuration as that of the embodiment of the second invention shown in FIG. 3, so its explanation is omitted and based on the operation flowchart shown in FIG. The operation will be described. FIG. 8 shows the CPU 1 of the voice recognition apparatus according to the third embodiment of the second invention.
6 is a flowchart showing the operation of FIG.

The channel selection signal CS supplied from the voice recognition unit 21 is monitored and a latch timing signal is output to the latch circuits 14 and 20 (step S41). The latch circuits 14 and 20 output the data M (CH) and R (C
H) is taken in (step S42).

Next, the voice section is detected from the fetched data (step S43). Start t of this voice section
The detection of s is performed as follows. Filter bank 13
Using the correction coefficient before updating (which is referred to as G0 (0)) obtained from the all-pass filter 13a, the level of the audio signal becomes higher than the predetermined threshold level TH by the following (Equation 6). Point (time), that is, starting point t
s is detected.

[0078]

In this (Equation 6), M (0) and R (0) are the main signal and the reference signal obtained from the all-pass filter 13a. As described above, the voice section T1
Is 2.6 seconds from the starting end ts.

Further, the stock data for the past 1 second from the starting end ts is updated. That is, the delay data MD (CH) and RD (CH) one second before is taken out from the RAM (step S44). This delay data is stocked and correction coefficient candidates are created (step S45).
That is, the correction coefficient calculation stock data is updated using the delay data MD (CH) and RD (CH) as the correction coefficient calculation data. Then, the cumulative values ΣMD (CH) and ΣRD (CH) of the delay data for the past one second are calculated, and the correction coefficient candidate Gc (CH) is obtained by (Equation 9).

[0080]

## EQU9 ## Next, it is determined whether or not the voice start edge is detected (step S46). If the voice start edge is detected, Gc
The correction coefficient is updated with (CH) = G (CH) (step S47).

FIG. 9 is a diagram showing how the voice start end of the voice recognition apparatus in this embodiment is detected. The correction coefficient is updated by detecting the voice start end. That is, the correction coefficient is updated by such a method from 2 seconds (point c) to 1 second (b) before the point a which is the voice start end in FIG.
The ratio of the cumulative value of the data of (point) becomes the correction coefficient. Therefore, since the output data is delay data, the correction coefficient is updated from the point b.

Therefore, even when the correction coefficient is updated for each utterance, the fluctuation of the residual component of the audio noise is reduced, and the waveform as shown in FIG. 6B is obtained. In the case of having the stationary noise removal function of, the residual component is likely to be removed as stationary noise.

In the case of this method, since the correction coefficient before updating is determined at the time of the previous utterance, the voice recognition device is voice activated by the utterance of a specific word such as "voice control", and then the control is performed. This is effective in the case of a system that recognizes words.

When it is not the voice start end in step S46, the audio noise component is subtracted (step S
48), updated correction coefficient G (CH) and delay data M
Using D (CH) and RD (CH), the data SD (CH) of the delayed audio signal is extracted by (Equation 8) and the data is output (step S49).

[0085]

## EQU00008 ## According to the third embodiment, the correction coefficient is updated at the voice start edge detection position, so that the false detection of the voice start edge is less likely to occur even when the correction coefficient is updated for each utterance.

Further, since the utterance interval of voice is usually 2 seconds or more, it is expected that voice data is not included from point c to point b shown in FIG. 9, and it is not necessary to judge the presence or absence of voice component. There are also advantages.

3. An embodiment of the third invention will be described.

A feature of the present invention is that the updated correction coefficient is further multiplied by a predetermined adjustment amount.

The block diagram of the voice recognition apparatus according to the first and second embodiments of the present invention described below has the same configuration as that of the second embodiment of the present invention shown in FIG. 3, and the description thereof will be omitted. .

FIG. 10 is a flow chart showing the operation of the CPU 15 of the voice recognition device in the embodiment of the third invention. The operation of the first embodiment of the third invention will be described with reference to this flowchart and FIG.

First, the channel selection signal CS from the voice recognition section 21 is monitored, the latch timing signal is supplied to the latch circuits 14 and 20 (step S51), and the timing of fetching from the latch circuits 14 and 20 to the CPU 15 is determined. Structure, data M (CH), R (CH)
Is taken in (step S52).

The voice section is detected from the fetched data (step S53). This detection is performed by the correction coefficient G0 (0) before update obtained from the all-pass filter 13a.
Is used as the start point of the voice by the data satisfying (Equation 6).

[0093]

## EQU00006 ## The voice section is set to 2.6 seconds (maximum voice section length of the voice recognition device) from this voice start end.

Next, data stock and delayed data are extracted (step S54). Update the stock data for the past few seconds (1 second in this case) from the present,
As a result, delay data MD (CH) and RD (CH) are obtained. Then, it is determined whether or not it is in the voice section (step S5
5) If it is not a voice section, delay data that does not include a delayed voice component is stocked to create a correction coefficient candidate (step S56). Specifically, when it is not in the voice section, the obtained delay data MD (CH), RD (CH) is used as the correction coefficient calculation data, and the correction coefficient calculation stock data is updated. Then, cumulative values ΣMD (CH), ΣRD (CH) for the past one second that does not include a voice component
And the correction coefficient candidate Gc (C
H) is calculated.

[0095]

## EQU9 ## After obtaining the correction coefficient candidate Gc (CH), 0.5
The correction coefficient is updated every second (step S57). That is, a counter is set, this counter is incremented when it is not in the voice section, and the correction coefficient candidate Gc (CH) is updated as the correction coefficient G (CH) every 0.5 seconds. Further, the updated correction coefficient is adjusted (step S
58). This adjustment uses the cumulative value ΣMD (CH) of the delay data MD (CH) that does not include the delayed voice component to obtain the adjustment amount α from the adjustment amount determination rule, and the correction coefficient of the all-pass filter (CH0) 13a is G ′ (0) (= G
(0) · α).

FIG. 11 is a diagram showing a rule for determining the adjustment amount of the correction coefficient in the voice recognition device of the third and fourth inventions described later. In FIG. 11, the horizontal axis represents the delay data MD (C
H) is the number of cumulative values ΣMD (CH), and the vertical axis is the adjustment amount α. Α is 1 until the number of accumulated values is 200, and 2
From 00 to 400, α becomes 1.3, and the value of α increases according to the cumulative value. As described above, the value of α in the adjustment amount determination rule shown in FIG. 11 is always 1 or more, and the subtraction amount is set to be larger in advance for the adjustment of the adjustment amount in the next process.

In the correction processing of the correction coefficient (step S59), when the subtraction result is negative when subtracting the audio component of CH0, an excessive subtraction amount flag indicating that the subtraction amount is too large is set, The cumulative value of the past constant time (in this case, 3 seconds) other than is calculated. For example, when 1 frame is set to 10 ms, it can be understood that if the cumulative value is 300, the subtraction is excessively complete. In such a case, it is possible to avoid oversubtraction by decrementing and adjusting the adjustment amount α.

In this case, the rule is that if the accumulated value is> 285, the adjustment amount is decremented. If the accumulated value is <250, the adjustment amount is incremented. Since the accumulated data is data for 3 seconds, this determination is also made every 3 seconds. .

If it is not in the voice section in step S55, subtraction processing of the audio noise component is performed (step S60). This subtraction process is performed with the updated correction coefficient G
Using (CH) and the delay data MD (CH) and RD (CH), the audio signal SD (CH) is extracted by (Equation 8), and the subtracted data is output (step S61).

The effect of this embodiment is that the correction coefficient can be updated according to the fluctuation of the audio noise level.

Usually, when the audio noise level becomes large, an estimation error occurs, so that a large amount of residual data of audio noise remains and the detection error of the voice section in the voice recognition unit increases. Therefore, by applying the invention of this embodiment, when the audio noise level is high, a large amount of subtraction is performed in the all-pass filter in accordance with the adjustment amount determination rule, so that the detection error of the voice section can be reduced.

On the other hand, when the audio noise level is low, the residual level of the audio noise is low, and if too much subtraction is performed, the voice section becomes narrow and the similarity decreases. In such a case, according to the adjustment amount determination rule, the degree of similarity can be increased by subtracting so that the all-pass filter is reduced.

Next, a second embodiment of the present invention will be described.

The feature of the second embodiment resides in that the updated correction coefficient is further multiplied by a predetermined adjustment amount, as in the first embodiment. However, this embodiment is different from the first embodiment in that the correction coefficient is not modified.

FIG. 12 is a flow chart showing the operation of the CPU 15 of the voice recognition device in the second embodiment of the third invention. The operation of the second embodiment will be described based on this flowchart and FIG.

First, the channel selection signal CS from the voice recognition section 21 is monitored, a latch timing signal is supplied to the latch circuits 14 and 20 (step S71), and the timing of fetching from the latch circuits 14 and 20 to the CPU 15 is determined. Structure, data M (CH), R (CH)
Is taken in (step S72).

The voice section is detected from the fetched data (step S73). This detection is performed by the correction coefficient G0 (0) before update obtained from the all-pass filter 13a.
Is used as the start point of the voice by the data satisfying (Equation 6).

[0108]

## EQU00006 ## The voice section is set to 2.6 seconds (maximum voice section length of the voice recognition device) from this voice start end.

Next, the data stock and the delayed data are taken out (step S74). Update the stock data for the past few seconds (1 second in this case) from the present,
As a result, delay data MD (CH) and RD (CH) are obtained. Then, it is determined whether or not it is in the voice section (step S7).
5) If it is not in the voice section, delay data that does not include the delayed voice component is stocked to create correction coefficient candidates (step S76). Specifically, when it is not in the voice section, the obtained delay data MD (CH), RD (CH) is used as the correction coefficient calculation data, and the correction coefficient calculation stock data is updated. Then, cumulative values ΣMD (CH), ΣRD (CH) for the past one second that does not include a voice component
And the correction coefficient candidate Gc (C
H) is calculated.

[0110]

## EQU9 ## After obtaining the correction coefficient candidate Gc (CH), 0.5
The correction coefficient is updated every second (step S77). That is, a counter is set, this counter is incremented when it is not in the voice section, and the correction coefficient candidate Gc (CH) is updated as the correction coefficient G (CH) every 0.5 seconds. Further, the updated correction coefficient is adjusted (step S
78). This adjustment uses the cumulative value ΣMD (CH) of the delay data MD (CH) that does not include the delayed voice component to obtain the adjustment amount α from the adjustment amount determination rule, and the correction coefficient of the all-pass filter (CH0) 13a is G ′ (0) (= G
(0) · α). The adjustment amount determination rule is as shown in FIG. 9 as in the fourth embodiment.

When it is not in the voice section in step S75, the subtraction process of the audio noise component is performed (step S79). This subtraction process is performed with the updated correction coefficient G
Using (CH) and the delay data MD (CH) and RD (CH), the audio signal SD (CH) is extracted by (Equation 8), and the subtracted data is output (step S80).

[0112]

## EQU00008 ## The effect of this embodiment is that the correction coefficient can be updated according to the fluctuation of the audio noise level.

Usually, when the audio noise level becomes high, an estimation error occurs, so that a large amount of residual data of audio noise remains, resulting in a large number of detection errors in the voice section in the voice recognition unit. Therefore, by applying the invention of this embodiment, when the audio noise level is high, a large amount of subtraction is performed in the all-pass filter in accordance with the adjustment amount determination rule, so that the detection error of the voice section can be reduced.

On the other hand, when the audio noise level is low, the residual level of the audio noise is low, and if too much is subtracted, the voice section is narrowed and the similarity decreases. In such a case, according to the adjustment amount determination rule, the degree of similarity can be increased by subtracting so that the all-pass filter is reduced.

4. An embodiment of the fourth invention will be described.

The feature of the present invention is that a known audio signal is adaptively used as a reference signal in order to prevent a reduction in the recognition rate of the speech recognition device in an audio noise environment, especially in a speech recognition device installed in a moving body such as a vehicle. In particular, it is to take measures against the superposition of vehicle running noise by a method of removing audio noise components.

The block diagram of the voice recognition apparatus of this embodiment also has the same configuration as that of the embodiment of the second invention shown in FIG. 3, and the description thereof will be omitted.

In FIG. 3, the main signal ma input from the microphone 11 is represented by the following (Equation 10).

[0119]

[Equation 10] Here, sa receives the voice from the generator, and the microphone 11
Is a voice signal output as an electric signal by the microphone 1
The conversion characteristic of 1 is added. Further, oa is an audio signal transmitted from the audio device 16.
Further, g is a transmission characteristic in which the audio signal oa is converted into sound by the conversion characteristic of the speaker 17, and the generated sound is propagated and reaches the microphone 11. Further, na is a running noise component of the vehicle.

Using data (M (CH), R (CH)) obtained by frequency-analyzing and digitizing the input signal (ma, oa), (Equation 10) can be expressed by the following (Equation 11).

[0121]

[Equation 11] However, in this (Equation 11), the left term and the right term are not always completely equal due to an error caused by digitizing the analog signal. In this formula, S (CH)
Is data obtained by digitizing the audio signal sa, and G (C
H) is a correction coefficient for multiplying R (CH) to predict the audio signal component S (CH) included in the main signal M (CH). N (CH) is data obtained by digitizing the traveling noise component na.

From this (Equation 11), the running voice component N
The audio signal S (CH) including (CH) is shown below (Equation 1)
It is represented by 2).

[0123]

[Equation 12] From this (Equation 12), the composite component of the audio signal S (CH) and the running noise component N (CH) included in the main signal M (CH) can be predicted.

This correction coefficient G (CH) is M (C) when no sound is generated and when the running noise is zero if the number of channels n, which is the frequency resolution, is large.
It can be estimated from the ratio of H) and R (CH). That is, if S (CH) = 0 and N (CH) = 0, then M (C
H) = R (CH) · G (CH), and the correction coefficient G (C
This is because H) can be represented as M (CH) / R (CH).

Here, it is important how accurately G (CH) can be estimated under the environment of fluctuating noise and steady noise not included in the reference signal.

In this case, assuming that the voice recognition unit 21 has a function of removing the stationary noise, the voice component may include traveling noise which is the stationary noise. Therefore, the data supplied to the speech recognition unit 21 is S. (CH) + N (CH) is sufficient. The method in which the voice recognition unit removes stationary noise is S. S method.

FIG. 13 is a flow chart showing the operation of the CPU 15 of the voice recognition device in the embodiment of the fourth invention. Based on this flowchart and FIG. 3, this fourth
The operation of the embodiment of the invention will be described.

First, the channel selection signal CS from the voice recognition section 21 is monitored, the latch timing signal is supplied to the latch circuits 14 and 20 (step S81), and the fetch timing from the latch circuits 14 and 20 to the CPU 15 is determined. Structure, data M (CH), R (CH)
Is taken in (step S82).

The voice section is detected from the fetched data (step S83). This detection is performed by the correction coefficient G0 (0) before update obtained from the all-pass filter 13a.
Is used as the start point of the voice by the data satisfying (Equation 6). The trigger level in this case is a fixed value.

[0130]

## EQU00006 ## The voice section is set to 2.6 seconds (maximum voice section length of the voice recognition device) from this voice start end.

Next, data stock and delayed data are taken out (step S84). Update the stock data for the past few seconds (1 second in this case) from the present,
As a result, delay data MD (CH) and RD (CH) are obtained. Then, it is determined whether or not it is a voice section (step S8).
5) If it is not in the voice section, the delay data not including the delayed voice component is stocked to create a correction coefficient candidate (step S86). Specifically, when it is not in the voice section, the obtained delay data MD (CH), RD (CH) is used as the correction coefficient calculation data, and the correction coefficient calculation stock data is updated. Then, cumulative values ΣMD (CH), ΣRD (CH) for the past one second that does not include a voice component
And the correction coefficient candidate Gc (C
H) is calculated.

[0132]

Next, various flags are set (step S8).
7). As the flags in this case, a flag (N-FLAG) indicating whether the vehicle is in a running noise environment and a flag (M-FLAG) indicating whether a vehicle is in a music noise environment are set.

The setting of N-FLAG is performed by accumulating the cumulative values ΣMD (1) and ΣRD (1) of the data not including the delayed voice component of CH1 obtained at step S86 and the cumulative value for 30 seconds not including the traveling voice component. Further, cumulative values ΣΣMD (1) and ΣΣRD (1) of ΣMD (1) and ΣRD (1) (which will be described later) are used, and N-FLAG is set when the following (Equation 13) is satisfied.

[0134]

[Equation 13] M-FLAG is set by the CH obtained in step S86.
The cumulative value ΣRD (0) of data that does not include the delayed audio component of the 0 reference signal is used, and M-FLAG is calculated when the following (Equation 14) is satisfied.

[0135]

[Equation 14] Hereinafter, the case where N-FLAG is accumulated is represented as NF = 1, the case where N-FLAG is not accumulated is represented as NF = 0, the case where M-FLAG is accumulated is represented as MF = 1, and the case where M-FLAG is not accumulated is represented as MF = 0.

After that, these two flags are judged,
Update of correction coefficient (step S88), adjustment of correction coefficient (step S89), correction of correction coefficient (step S9)
0) is performed. For updating the correction coefficient, NF = 0, M
When F = 1, the counter is set, the counter is incremented when it is not in the voice section, and the correction coefficient is updated as G (CH) = Gc (CH) at regular time intervals (0.5 seconds in this case). To do.

Further, the 30-second average correction coefficient is calculated. That is, the accumulated values ΣMD (CH) and ΣRD (C of the data not including the delayed voice component obtained in step S86
H) is stocked at regular intervals (0.5 seconds in this case), and at the same time, a further cumulative value ΣΣMD for the past 30 seconds.
(CH) and ΣΣRD (CH) are obtained. C obtained here
The cumulative value of H0 becomes the data used in (Equation 13).

When NF = 1 and MF = 1, the correction coefficient is determined by the following (Equation 15).

[0139]

[Equation 15] The adjustment of the correction coefficient is also performed differently depending on the values of the two flags. When NF = 0 and MF = 1, the cumulative value ΣMD (0) of the data that does not include the delayed audio component of the main signal is used. Then, the adjustment amount α is obtained from the adjustment amount determination rule of FIG. 9, and the all-pass filter (CH0) 13
Let the correction coefficient of a be G '(0) (= G (0) · α).
Here, the adjustment amount determination rule is set in advance so that the subtraction amount is increased in order to correct the next adjustment amount.

On the other hand, when NF = 1 and MF = 1, the cumulative value ΣM of data not including the delayed audio component of the main signal
Since D (0) contains a running noise component, the following (Equation 1)
The estimated value ΣMD inf (0) of ΣMD (0) is obtained by 6).

[0141]

[Equation 16] Then, as in the case of NF = 0 and MF = 1, the adjustment amount α
From the adjustment amount determination rule of FIG. 9, and the correction coefficient of the all-pass filter (CH0) 13a is G ′ (0) (= G
(0) · α).

Next, regarding the correction of the correction coefficient, MF
Focusing only on LAG, when MF = 1, the adjustment amount is corrected. When the subtraction result becomes negative when subtracting the audio component of CH0, an excessive subtraction amount flag is set to indicate that the subtraction amount is too large, and the cumulative past past time (3 seconds in this case) other than the voice section is accumulated. Calculate the value. For example, when 1 frame is set to 10 ms, it can be understood that if the cumulative value is 300, the subtraction is excessively complete. In such a case, it is possible to avoid oversubtraction by decrementing and adjusting the adjustment amount α.

The rule in this case is that if the cumulative value> 285, the adjustment amount is decremented. If the cumulative value <250, the adjustment amount is incremented. Since the cumulative data is 3 seconds data, this judgment is also made every 3 seconds. .

If it is not in the voice section in step S85, a flag is set (step S91) and an audio noise component subtraction process is performed (step S92).
In this subtraction process, the audio signal SD (CH) is extracted by (Equation 8) using the updated correction coefficient G (CH) and the delay data MD (CH) and RD (CH), and the subtracted data is extracted. Output (step S93).

According to the fourth embodiment of the present invention, the adaptive S.V. In the S method, even when the running noise is superimposed on the voice component, the audio noise component is removed with an appropriate correction coefficient, and the function of the voice recognition unit can be used as it is for the running noise removal.

FIG. 27 is a diagram showing how a voice start edge of the voice recognition device in the third embodiment of the second invention is detected.

5. Embodiments of the fifth to seventh inventions will be described.

The fifth to seventh inventions are a fuzzy method for determining the voice trigger level, a method for determining the running noise determination level, and a method for determining the adjustment amount of the correction coefficient, respectively, in the first to fourth inventions. It is based on inference.

Since the system configurations of these embodiments are the same as those in the block diagram of FIG. 3, the description thereof will be omitted.
FIG. 14 is a method for determining a voice trigger level, a method for determining a running noise determination level according to the fifth to seventh inventions, and
9 is a flowchart showing the operation of the CPU 15 when the method of determining the adjustment amount of the correction coefficient is performed by fuzzy inference.

In FIG. 3 and FIG. 14, the voice recognition unit 21
Monitor the CH signal and then latch circuits 14 and 2
The latch timing of 0 and the fetch timing of the CPU 15 are created to fetch the data (M (CH), R (CH)) (step S101). Next, the voice trigger level is determined by fuzzy inference to detect the voice section (step S102). In this case, the correction coefficient before update is used as G0 (CH),

[0151]

Let [Equation 6] be the beginning of the voice. The voice section is 2.6 seconds from the beginning of the voice (this is the maximum voice section length of the voice recognition device). In this case, similarly to the second to fourth inventions, stock data from the voice start end, that is, from the present time to the past several seconds (in this case, one second) is read from the RAM and delay data MD (CH) and RD (CH) are obtained. obtain. Then, it is determined whether or not it is a voice section (step S103). If it is not a voice section, delay data MD (CH), RD (CH)
The stock data for calculating the correction coefficient is updated by using as the calculation data for the correction coefficient. Then, ΣMD (CH) and ΣRD, which are the cumulative values of the delay data not containing the voice component in the past 1 second.
(CH) and calculate the correction coefficient candidate Gc (CH) as

[0152]

Created from

Then, it is judged whether or not there is an acoustic noise, that is, an audio noise component (step S104). A flag indicating whether or not the environment is an acoustic noise environment (this is referred to as "M FLAG") is set according to the determination result.
This setting uses the cumulative value ΣRD (1) of the delayed voice component-free data of the CH0 reference signal,

[0154]

When the condition of [Equation 14] is satisfied, M FLAG is set.

When there is an audio noise component, it is determined by fuzzy inference whether or not there is running noise (step S105). A flag indicating whether or not the vehicle is in a traveling noise environment (this is referred to as "N FLAG") is set according to the determination result. In this setting, the running noise determination level is determined by fuzzy inference (step S10).
5), at this time, the acoustic noise level ΣMD obtained last time
Determine using (0). Then, the cumulative values ΣMD (1) and ΣRD (1) of the delayed voice component-free data of CH1.
And a cumulative value for 30 seconds that does not include a running noise component (this will be described later),

[0156]

When the condition of [Equation 13] is satisfied, N FLAG is set.

M FLAG is on, N FLAG
If not, the correction coefficient is adaptively updated (step S106). Specifically, a counter is set, this counter is advanced when it is not in the voice section, and G (CH) = G every few milliseconds (0.5 seconds in this case).
The correction coefficient is updated as c (CH). Further, the cumulative value (ΣMD (CH), ΣRD (CH)) of the delayed voice component-free data obtained in step S103 is stocked every few seconds (here, 0.5 seconds), and at the same time, the further 30 seconds have passed. Cumulative values (ΣΣMD (CH), ΣΣRD (CH)) are obtained. Then, the adjustment amount is determined and corrected by fuzzy estimation (step S107), and the subtraction process is further performed by the determined or corrected adjustment amount (step S10).
8) and outputs the subtraction result (step S109).

On the other hand, when both M FLAG and N FLAG are present, ΣΣMD (CH) and ΣΣRD (CH) which are the cumulative values of the delayed voice component non-containing data obtained in step S106 for the past 30 seconds are used. , The following equation,

The correction coefficient is updated by the following equation (15) (step S11).
0). The parameter is estimated by the updated correction coefficient (step S111), the adjustment amount is determined and corrected by fuzzy reasoning (step S107), and the subtraction process is performed by the determined or corrected adjustment amount (step S1).
08), and outputs the subtraction result (step S10).
9). If it is a voice section in step S103 and if there is no audio noise component in step S104, the process proceeds to step S108 without performing the adjustment amount determination and correction, and the subtraction process is performed.

A method for determining a voice trigger level according to the fifth aspect of the invention will be described. FIG. 15 shows a fuzzy inference method for determining the voice trigger level in step S102. That is, in the embodiment of the fourth invention, the voice section is detected by the fixed value voice trigger level, but in the present invention, the voice trigger level is determined based on the fuzzy rule of FIG.

Adaptive S. S. Since the method determines the voice start edge based on the result of preprocessing subtraction, refer to the result of preprocessing subtraction to determine the fuzzy rule for determining the audio trigger level of this method. Created. This fuzzy rule is a method based on the MAX-MIN centroid method. The centroid method is a method of obtaining an inference result for each rule (in this case, rules 1 to 6), combining the inference results in each rule, and obtaining the inference result of the entire rule as its centroid. FIG. 16 shows a result of subtraction corresponding to the fuzzy rule shown in FIG.
It is assumed that 6 is shown in FIGS. 16 (a) to 16 (f).
That is, the voice trigger level is adjusted according to the residual level.

The following effects can be obtained by the method for determining a voice trigger level by fuzzy inference according to the present invention.

1. It is possible to determine the voice trigger level according to the acoustic noise level and the running noise level. When the residual component is large, increase the trigger level so that the residual component will not trigger, and when the residual component is small, compare The adaptive S.I.S. S. It has become possible to improve the effect of the method.

2. By using fuzzy reasoning,
It has become easy to create and adjust a control rule based on a plurality of parameters (here, two parameters of the acoustic noise level and the running noise level) that were difficult to achieve by normal control.

3. The fuzzy rule of the invention is shown by six rules, but due to the interpolation effect by fuzzy inference,
Appropriate control is possible even for intermediate values, and fine control is possible.

4. Considering the Lombard effect, in the state where the noise level is low (as a result, when the residual level of noise according to this method is small, for example, FIG. 16D), the utterance level is low even for the same speaker. In the opposite case, the Lombard effect raises the vocalization level. Therefore, this method matches the fluctuation of the voice level due to the Lombard effect.

Next, a traveling noise determination level determining method according to the sixth aspect of the present invention will be described. In FIG. 17, for determining the presence or absence of running noise in step S105 of FIG.
The threshold value determination rule of a running noise determination flag is shown. Calculation of the inference result of the entire rule is based on the MAX-MIN centroid method.

In the flag setting in FIG. 13, the threshold value for determining the running noise and setting the flag is fixed. As this fixed value, when the acoustic noise level is small to some extent, the traveling noise level before the traveling noise level at which the traveling noise becomes dominant over the acoustic noise is adopted. However, in reality, when the acoustic noise level is considerably high, the traveling noise level in which the traveling noise becomes dominant over the acoustic noise shifts upward.

Therefore, the correction coefficient is not updated and learned at a position where the traveling noise level is lower than necessary, and the effect of adaptive processing is reduced. Therefore, the threshold value of the running noise determination flag is determined according to the sound level by fuzzy inference.

This rule considers which of acoustic noise and running noise is dominant. In other words, when the acoustic noise level is “substantially high”, the acoustic noise generally tends to be dominant, and the traveling noise level at which the traveling noise is dominant is located considerably above. Therefore, the threshold value of the traveling noise determination flag is also “ It's quite large. "

On the other hand, when the acoustic noise level is "small to some extent", the traveling noise is likely to be dominant and the traveling noise level is located at a low level, so the threshold value of the traveling noise judgment flag is "small". And

According to the sixth aspect of the present invention, it is possible to determine the threshold value of the running noise determination flag according to the acoustic noise level by fuzzy inference, and the adaptive S.I. S. The effect of the method can be improved.

Further, although the adjustment amount determination rule of the present invention is shown by two rules according to the acoustic noise level, the interpolation effect by the fuzzy reasoning makes it possible to appropriately control even an intermediate value, and to perform fine adjustment. Control became possible.

Next, a method of determining an adjustment amount by fuzzy inference according to the seventh invention will be described. FIG. 18 shows an adjustment amount determination rule for determining the adjustment amount in step S108 of FIG. Further, FIG. 19 is a diagram showing a subtraction result when the subtraction amount is changed according to the traveling noise level. In FIG. 19 (a), the traveling noise level is "very high".
19 shows the subtraction result in the case of the normal subtraction amount at
FIG. 19B shows the subtraction result when the subtraction amount is small when the traveling noise level is “somewhat high”, and FIG.
Indicates the subtraction result when the amount of subtraction is increased when the running noise level is “somewhat high”. Further, FIG. 20 is a diagram schematically showing the relationship between the traveling noise level and the acoustic noise level and the adjustment amount. In FIG. 20, a is based on the rules 1 and 2 in FIG. 18 when the running noise level is “small”, b is based on the rules 1 and 3 when the running noise level is “somewhat high”, and c Is due to Rule 1 and Rule 4 when the running noise level is "significantly high".

The method of the adjustment amount determination rule in FIG. 18 is a further improvement of the adjustment amount determination rule in the third and fourth inventions. This fuzzy rule is MAX-
It is a method based on the MIN centroid method.

Rule 1 is that regardless of the running noise level,
When the acoustic noise level is "quite low", the adjustment amount is set to "low".

Rule 2 sets the adjustment amount to "high" when the traveling noise level is "low" and when the acoustic noise level is "high".

Rule 3 sets the adjustment amount to "substantially large" when the traveling level is "somewhat large" and the acoustic noise level is "large".

Rule 4 sets the adjustment amount to "large" when the running noise level is "quite large" and the acoustic noise level is "large".

This fuzzy rule is created with the expectation of the following effects.

1. In an environment where the running noise is “somewhat loud” but not so dominant to the acoustic noise, the noise component is subtracted to a large extent by setting the adjustment amount to a larger value than when the running noise level is “small”. Is more effective.
This is because the acoustic noise component is likely to remain when the traveling noise is superimposed with a large estimation error of the noise component, so that it is less subtracted to leave the traveling noise component + the residual component of the acoustic noise (FIG. 19B).
This is because it is more difficult to trigger the voice trigger in the recognition unit when the residual component of the traveling noise is left by subtracting the traveling noise component and subtracting the traveling noise component, as shown in FIG. 20-c.

2. When the driving noise is "quite loud",
In an environment in which the acoustic noise is dominant over the acoustic noise, the acoustic noise is buried in the traveling noise (see FIG. 19A), and thus the subtraction amount may be small.

After all, the relationship between the acoustic noise level and the adjustment amount according to the running noise level is set as shown in FIG.
The above effects can be expected.

[0183]

As is apparent from the above embodiments, the following effects can be obtained by the first to seventh inventions.

1. EFFECTS OF THE FIRST INVENTION According to the present invention, by always obtaining the latest correction coefficient from the main signal and the reference signal in the non-voice section, it is possible to deal with non-stationary noise such as abruptly changing audio noise. If the audio noise does not change so much, the convergence time of the filter can be shortened. Further, there is no need for an expensive processing device capable of high-speed arithmetic processing such as DSP.

Further, since the reference signal does not include the voice signal, the estimation error can be reduced and the high voice recognition can be performed even in the audio noise environment.

2. Effects of the Second Invention The effects of the second invention are as follows.

1) Even if the audio noise is large, the audio noise component is removed to some extent in advance by the subtraction data using the pre-update correction coefficient, so that the detection error at the beginning of the audio signal can be reduced.

2) Since the voice section is automatically detected, it is possible to eliminate the burden of the user performing an operation such as key input each time the user speaks.

3) Even if there is a residual component of audio noise component due to an estimation error because the level of the voice signal is small, the threshold level for detecting the voice signal is set to a large value, thereby erroneously detecting the voice section due to the audio noise. Can be reduced and an appropriate value of the correction coefficient can be obtained. Therefore, it does not depend on the threshold level of the voice recognition unit.

4) Real-time processing is possible because of the extremely simple method.

3. Effect of the third invention The effect of the present invention is that the correction coefficient can be updated according to the fluctuation of the audio noise level.

Usually, when the audio noise level becomes large, an estimation error occurs, so that a large amount of residual data of the audio noise remains and the detection error of the voice section in the voice recognition unit increases. Therefore, by applying the invention of this embodiment, when the audio noise level is high, a large amount of subtraction is performed in the all-pass filter in accordance with the adjustment amount determination rule, so that the detection error of the voice section can be reduced.

On the other hand, when the audio noise level is low, the residual level of the audio noise is low, and if too much subtraction is performed, the voice section becomes narrow and the similarity decreases. In such a case, according to the adjustment amount determination rule, the degree of similarity can be increased by subtracting so that the all-pass filter is reduced.

4. EFFECTS OF FOURTH INVENTION According to the present invention, in the adaptive spectral subtraction method under the known noise environment, even when the running noise is superposed on the voice component, the audio noise component is removed with the appropriate correction coefficient. The function of the voice recognition unit can be used as it is for removing the traveling noise.

5. Effects of Fifth Invention According to the fifth invention, there are the following effects.

1) The fuzzy inference of the present invention makes it possible to determine the voice trigger level according to the acoustic noise level and the running noise level, and when the residual component is large, the residual component does not trigger. By increasing the trigger level and reducing the trigger level so as not to make it difficult for the utterer having a relatively low voicing level to easily trigger the voicing sound when the residual component is small, the adaptive S.I. S. The effect of the method could be improved.

2) By using fuzzy inference,
It has become easy to create and adjust a control rule based on a plurality of parameters (here, two parameters of the acoustic noise level and the running noise level) that were difficult to achieve by normal control.

3) The fuzzy rule of the present invention is shown by six rules, but due to the interpolation effect by fuzzy inference,
Appropriate control is possible even for intermediate values, and fine control is possible.

4) Considering the Lombard effect, when the noise level is low (as a result, the residual level of noise according to this method is small (for example, FIG. 2D)), the utterance level is low even for the same speaker. In the opposite case, the Lombard effect raises the vocalization level. Therefore, this method matches the fluctuation of the voice level due to the Lombard effect.

6. Effects of Sixth Invention According to the sixth invention, there are the following effects.

1) By the fuzzy inference of the present invention, it becomes possible to determine the threshold value of the running noise determination flag according to the acoustic noise level, which is more adaptive than the conventional system. S.
The effect of the method was able to be improved.

2) The adjustment amount determination rule of the present invention is shown by two rules according to the acoustic noise level, but due to the interpolation effect by fuzzy reasoning, appropriate control is possible even for intermediate values, and fine adjustment is possible. Control became possible.

7. Effects of Seventh Invention According to the seventh invention, there are the following effects.

1) By fuzzy reasoning, it becomes possible to determine the adjustment amount according to the acoustic noise level and the running noise level. When there is running noise to some extent, the adjustment amount is made larger than usual and the subtraction amount is increased. , When the driving noise is large, the adjustment amount is made smaller than usual to reduce the adjustment amount to a smaller amount, so that the adaptive S.S. S. The effect of the method was able to be improved.

2) By using fuzzy inference,
It has become easy to create and adjust a control rule based on a plurality of parameters (here, two parameters of the acoustic noise level and the running noise level) that were difficult to achieve by normal control.

3) The adjustment amount determination rule of the present invention is shown by three rules when the acoustic noise level is "large". However, due to the interpolation effect by fuzzy reasoning, appropriate control is performed even for intermediate values. It became possible, and fine control became possible.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device in an embodiment of the first invention.

FIG. 2 is a flowchart showing an operation of a CPU 15 of the voice recognition device shown in FIG.

FIG. 3 is a block diagram of a voice recognition device in a first embodiment of the second invention.

FIG. 4 is a diagram showing how a voice start end of the voice recognition device shown in FIG. 3 is detected.

FIG. 5 is a flowchart showing an operation of a CPU 15 of the voice recognition device in the first exemplary embodiment of the second invention.

FIG. 6A is a diagram showing a residual amount of an audio component when a correction coefficient is updated every time a voice is uttered. (B)
[Fig. 6] is a diagram showing a residual amount of audio components when a correction coefficient is updated at regular intervals.

FIG. 7 is a flowchart showing an operation of the voice recognition device in the second exemplary embodiment of the second invention.

FIG. 8 is a flowchart showing an operation of a CPU 15 of the voice recognition device in the third exemplary embodiment of the second invention.

[Figure 9]

FIG. 10 is a flowchart showing an operation of the CPU 15 of the voice recognition device in the first exemplary embodiment of the third invention.

FIG. 11 is a diagram showing an adjustment amount determination rule of a correction coefficient in the voice recognition device of the third and fourth inventions.

FIG. 12 is a flowchart showing the operation of the CPU 15 of the voice recognition device in the second exemplary embodiment of the third invention.

FIG. 13 is a flowchart showing an operation of the CPU 15 of the voice recognition device in the embodiment of the fourth invention.

FIG. 14 shows the operation of the CPU 15 when the method of determining the voice trigger level, the method of determining the running noise determination level, and the method of determining the adjustment amount of the correction coefficient in the fifth to seventh inventions are performed by fuzzy inference. It is a flowchart showing.

15 is a diagram showing a fuzzy inference method for determining a voice trigger level in step S102 of FIG.

16 is a diagram showing a result of subtraction corresponding to the fuzzy rule of FIG.

FIG. 17 is a diagram showing a threshold value determination rule of a traveling noise determination flag for determining the presence or absence of traveling noise in step S105 of FIG.

FIG. 18 is a diagram showing an adjustment amount determination rule for determining an adjustment amount in step S108 of FIG.

FIG. 19 is a diagram showing a subtraction result when the subtraction amount is changed according to the traveling noise level.

FIG. 20 is a diagram showing an outline of a relationship between a traveling noise level and an acoustic noise level and an adjustment amount.

FIG. 21 is a block diagram of a conventional voice recognition device.

[Explanation of symbols]

 Reference Signs List 11 microphone 13 filter bank 15 CPU 16 audio device 18 amplifier 19 filter bank 21 voice recognition unit (recognition means) 22 registration dictionary

Claims (7)

[Claims] 1. A voice recognition device for recognizing a voice signal by removing the noise component from a main signal in which a voice signal from a speaker is mixed with a noise component and comparing the voice signal with a comparative voice signal registered in advance. A means for generating a reference signal based on the noise component, a voice section in which the voice signal is included in the main signal,
A voice section discriminating means for discriminating a non-voice section which is not included; a correction coefficient updating means for generating and updating a correction coefficient based on the main signal in the non-voice section; A calculation means for subtracting a value multiplied by a correction coefficient from the main signal; and a recognition means for performing voice recognition by collating the calculation result obtained from the calculation means with the comparative voice signal. Voice recognition device. 2. A voice start edge detecting means for setting a temporary voice start edge of the voice signal when a subtraction result obtained by subtracting a value obtained by multiplying the reference signal by a correction coefficient before updating from the main signal is larger than a predetermined value. A means for generating a delayed main signal and a delayed reference signal from a past main signal and a reference signal before a fixed time, and a voice section determination means for determining a fixed voice start end based on the delayed main signal and determining the voice section A correction coefficient updating means for generating and updating a correction coefficient from a ratio of cumulative values of the delayed main signal and the delayed reference signal in a certain past time; and a value obtained by multiplying the delayed reference signal by the correction coefficient in the voice section. The speech recognition apparatus according to claim 1, further comprising: an arithmetic unit that subtracts from the delayed main signal. 3. A means for generating a correction coefficient candidate from a ratio of cumulative values of the delayed main signal and the delay reference signal in the past fixed time, a correction coefficient updating means for updating the correction coefficient candidate at fixed time intervals, and a delay. A correction coefficient adjusting unit that adjusts the correction coefficient based on a predetermined adjustment amount determination rule by using a delayed main signal that does not include an audio signal, and a correction coefficient correcting unit that corrects the adjusted correction coefficient are provided. The voice recognition device according to claim 2, wherein 4. A means for generating a correction coefficient candidate from a ratio of cumulative values of the delayed main signal and the delayed reference signal for a certain past time, and whether or not the noise component includes musical noise and the noise component is a vehicle. Noise component discriminating means for discriminating whether or not the traveling noise is included, and when the noise component includes only the music noise, the value of the correction factor candidate is updated as a correction factor at fixed time intervals, and the noise component When the includes the music noise and the running noise, a correction coefficient updating unit that uses a ratio of a cumulative value of the cumulative values of the delayed main signal and the delayed reference signal for a certain past time as a correction coefficient, and the noise component is the musical noise. When including only, the correction coefficient is adjusted based on a predetermined adjustment amount determination rule, and when the noise component includes the music noise and running noise,
Correction coefficient adjusting means for adjusting the correction coefficient based on the predetermined adjustment amount determination rule using the estimated main signal obtained from the calculating means; and when the noise component includes the music noise, the adjustment is made. 3. The voice recognition device according to claim 2, further comprising a correction coefficient correction unit that corrects the correction coefficient. 5. A voice recognition method in a noise environment for recognizing a voice of the speaker by removing the noise component from an input signal in which a noise component is mixed with a voice signal component from the speaker, the input signal From this, a voice section is detected by fuzzy inference, it is determined whether or not the noise component is mixed in this voice section, and the correction count for predicting the voice signal component is updated according to the determination result, A voice recognition method in a noise environment, characterized in that the updated correction count is adjusted, subtraction processing is performed based on the adjusted correction count, and voice recognition is performed using the subtraction result as the voice signal component. 6. A voice recognition method in a noise environment for recognizing a voice of the speaker by removing the noise component from an input signal in which an acoustic noise component and a running noise component are mixed in a voice signal component from the speaker. A correction count for detecting a voice section from the input signal, determining by fuzzy inference whether or not the running noise component is mixed in the voice section, and predicting the voice signal component according to the determination result. In the noise environment characterized by performing the adjustment of the updated correction count, performing subtraction processing based on the adjusted correction count, and performing voice recognition using the subtraction result as the voice signal component. Voice recognition method. 7. A voice recognition method for recognizing a voice of the speaker by removing the noise component from an input signal in which a noise component is mixed with a voice signal component from the speaker, wherein a voice section is recognized from the input signal. Detecting and determining whether or not the noise component is mixed in this voice section, updating the correction count predicting the voice signal component according to the determination result, and updating the correction count by fuzzy inference. Is performed, subtraction processing is performed based on the adjusted correction count, and voice recognition is performed using the subtraction result as the voice signal component.