JP5457293B2 - Voice recognition device - Google Patents

Description

  The present invention relates to a technique for controlling an input gain of a voice signal in a voice recognition device.

  As a technique for controlling the input gain of a speech signal in the speech recognition apparatus, a technique for adjusting the input gain according to a past speech recognition success rate (Patent Document 1), or a time interval targeted for recognition in past speech recognition There is known a technique (Patent Document 2) for setting an input gain in accordance with an audio signal level.

Japanese Patent No. 2975808 Japanese Patent No. 3594356

For example, when the voice recognition device is used in a situation where the surrounding acoustic environment such as in an automobile changes every moment, the noise level included in the voice signal also changes every moment.
Therefore, as described above, if the input gain is adjusted based only on the situation at the time of past speech recognition, such as the success rate of past speech recognition and the speech signal level in the past speech section, An input gain suitable for the noise situation cannot be set.
Therefore, an object of the present invention is to provide a speech recognition apparatus that can set an input gain more suitable for the current noise situation as an input gain of a speech signal.

  In order to achieve the above object, the present invention provides a voice recognition device that performs voice recognition, a microphone, an input amplifier that amplifies an input voice signal output from the microphone, and a signal amplified by the input amplifier. An AD converter for converting data, a speech recognition engine for performing speech recognition processing on input speech data output from the AD converter in response to a speech recognition execution instruction, a noise level detection unit, and speech speech A level detection unit and an input gain control unit for controlling the gain of the input amplifier are included. Here, in the voice recognition engine, in the voice recognition process, a time interval in which the input voice data includes the user's uttered voice is detected as the uttered voice period, and is included in the input voice data of the detected uttered voice period. Identifying the content of the uttered voice, the noise level detection unit, in a time section other than the utterance voice section, or a time section in which the voice recognition processing is not performed, the noise level included in the input voice signal, Repetitively calculating based on the input voice data, the utterance voice level detection unit, the average level of the utterance voice included in the input voice signal of each utterance voice section detected in each time of the voice recognition processing, The input gain control unit calculates the noise level at the start of each speech recognition process. The input of the speech speech section detected by the speech recognition process of the current time based on the level of noise finally calculated by the detection unit and the average level of speech speech detected by the speech speech level detection unit Estimating the level of the voice signal, and setting the gain of the input amplifier so that the level obtained by amplifying the estimated level of the input voice signal by the input amplifier is a level suitable for the voice recognition engine. .

  According to such a voice recognition device, the noise level detection unit repeatedly executes the noise level, and at the start of the voice recognition process, the noise level detected last, and therefore the noise level at the most recent time point, The level of the input voice signal is estimated based on the average level of the uttered voice at the time of the previous voice recognition processing execution, and the level obtained by amplifying the estimated level of the input voice signal by the input amplifier is the voice recognition. Set the gain of the input amplifier to a level suitable for the engine.

  The noise level at the most recent time can be expected to approximate the noise level in the current noise situation. Therefore, according to such a speech recognition apparatus, a gain more suitable for the current noise situation can be set in the input amplifier at the start of speech recognition processing.

  Here, in the speech recognition apparatus as described above, in the input gain control unit, during the time interval in which the speech recognition processing is not performed, the gain of the input amplifier is set to the maximum level that the level of the input speech signal can take. Is set to a predetermined value so that the level amplified by the input amplifier does not exceed the input range of the AD converter, or in a time interval in which the speech recognition processing is not performed This is preferable in ensuring proper calculation of the noise level.

  Further, when the above voice recognition device is used together with an audio device that outputs an audio sound represented by audio data, the audio recognition device of the audio device is in the period during which the voice recognition processing is performed. An output suppression unit that suppresses the output of the input audio signal and the audio data in the noise level detection unit in a time interval in which the speech recognition process is not performed. You may make it calculate based on these.

  More specifically, in the above speech recognition apparatus, the noise level detection unit calculates a noise amplitude distribution as the noise level, and the utterance speech level detection unit calculates the average of the utterance speech. Alternatively, an average amplitude distribution of the uttered voice may be calculated as the level, and the input gain control unit may estimate the amplitude distribution of the input voice signal as the level of the input voice signal.

  In this case, when the dynamic range of the amplitude distribution range indicated by the estimated amplitude distribution of the input speech signal is less than or equal to the dynamic range of the input range of the speech recognition engine in the input gain control unit. The amplitude value after the amplitude value at the center of the amplitude distribution range indicated by the estimated amplitude distribution of the input speech signal is amplified by the input amplifier becomes the amplitude value at the center of the input range of the speech recognition engine. It is preferable to set the gain of the input amplifier.

  In this case, in the input gain control unit, if the dynamic range of the amplitude distribution range in the estimated amplitude distribution of the input speech signal exceeds the dynamic range of the input range of the speech recognition engine, Of the amplitude distribution range in the estimated amplitude distribution of the input speech signal, the range portion having the same dynamic range as the dynamic range of the input range of the speech recognition engine, and the sum of the frequencies in the range distribution is the maximum Preferably, the gain of the input amplifier is set so that the range after the selected partial range is amplified by the input amplifier matches the input range of the speech recognition engine.

  As described above, according to the present invention, it is possible to provide a speech recognition apparatus that can set an input gain more suitable for the current noise situation as an input gain of a speech signal.

It is a block diagram which shows the structure of the speech recognition system which concerns on embodiment of this invention. It is a timing chart which shows operation | movement of the speech recognition system which concerns on embodiment of this invention. It is a flowchart which shows the input gain control process which concerns on embodiment of this invention. It is a figure which shows the process example of the input gain control process which concerns on embodiment of this invention.

Hereinafter, embodiments of the present invention will be described.
FIG. 1 shows the configuration of a speech recognition system according to this embodiment.
As shown in the figure, the speech recognition system includes a DA converter 1 that DA converts audio data output from an audio device (not shown) into an analog audio signal, an output amplifier 2 that amplifies the audio signal with an output gain SpG, and an audio amplifier. The sound represented by the output audio signal is output as the speaker output sound. The speaker 3, the microphone 4, the input sound signal representing the sound picked up by the microphone 4, the input amplifier 5 that amplifies the sound by the input gain G, and the input that is amplified by the input amplifier 5 An AD converter 6 for digitally converting a voice signal into input voice data, a voice recognition engine 7 for executing voice recognition processing on the input voice data converted by the AD converter 6, a talk switch 8, and an output gain of the output amplifier 2 The output gain control unit 9 for controlling SpG and the gain G of the input amplifier 5 are controlled. And an input gain control unit 10.

  In such a configuration, the voice recognition engine 7 starts the voice recognition process when the user depresses the talk switch 8. In the voice recognition processing, detection of an utterance voice section, which is a section in which the user's utterance voice is included in the input voice data, and voice recognition (identification of user utterance contents) for the input voice data in the utterance voice section are performed. The speech recognition engine 7 outputs a speech recognition in-progress signal Ron that is turned on from the time the user depresses the talk switch 8 to the end of the speech speech section, and also detects during the speech recognition processing after the speech recognition processing ends. The utterance voice section data Son representing the time position of the uttered voice section is output.

  The output gain control unit 9 controls the output gain SpG of the output amplifier 2 according to the volume signal Vol output from the audio device during the period in which the voice recognition in-progress signal Ron is off, and the voice recognition in-progress signal Ron is on. During this period, the output gain SpG of the output amplifier 2 is set to 0 to suppress the generation of speaker output sound.

Here, the input audio signal output from the microphone 4 includes, as its components, the speaker output sound a, noise b, and the user's uttered speech s output from the speaker 3.
The input gain control unit 10 includes a first noise amplitude distribution detection unit 11 that calculates the amplitude distribution of the noise b, a second noise amplitude distribution detection unit 12 that calculates the amplitude distribution gb (n) of the noise b, and noise. b, the noise amplitude distribution register 13 for storing the latest amplitude distribution gb (n), the voice amplitude distribution detecting unit 14 for detecting the average amplitude distribution f (n) of the uttered voice s, and the average voice of the uttered voice s. A voice amplitude distribution register 15 that stores the amplitude distribution f (n), a convolution calculator 16, and a gain control unit 17 are provided. Note that n in the amplitude distribution Z (n) represents that the amplitude distribution Z (n) represents the amplitude distribution by discretizing the amplitude value (dB) into a class of n amplitude values. .

Here, the timing of calculation of the amplitude distribution of the first noise amplitude distribution detection unit 11, the second noise amplitude distribution detection unit 12, and the audio amplitude distribution detection unit 14 will be described with reference to FIG.
Here, in FIG. 2, the input voice signal output from the microphone 4 is represented as x, and the speech voice section represented by the speech voice section data Son output from the voice recognition engine 7 is represented as SonD.
As shown in the figure, during the time interval until the talk switch 8 is depressed, the input audio signal x output from the microphone 4 includes a speaker output sound a and noise b as its components.
The first noise amplitude distribution detecting unit 11 is a time interval in which the speaker output sound a and the noise b are included as components in the input sound signal x, and a period during which the sound recognition signal Ron is off is set as a calculation execution period. During the calculation execution period, the amplitude distribution gb (n) of the noise b is calculated. Here, the details of the calculation method of the amplitude distribution gb (n) of the noise b in the first noise amplitude distribution detector 11 will be described later.

  Next, when the talk switch 8 is pushed down and the voice recognition signal Ron is turned on, the speaker output sound is suppressed. Therefore, after the voice recognition signal Ron is turned on, the voice recognition signal Ron is turned off. During the period other than the utterance voice section SonD represented by the utterance voice section Son, the input voice signal x output from the microphone 4 includes only noise b as its component. .

  The second noise amplitude distribution detection unit 12 is represented by the speech voice section data Son during the period in which the voice recognition signal Ron is on, which is a time section in which only the noise b is included as a component in the input voice signal x. The amplitude distribution gb (n) of the noise b is calculated during the calculation execution period, with the period other than the speech voice section SonD as the calculation execution period. Here, details of a method of calculating the amplitude distribution gb (n) of the noise b of the second noise amplitude distribution detection unit 12 will be described later.

Next, since the speech voice section SonD is a time section in which the user is speaking, the input voice signal x output from the microphone 4 includes noise b and speech voice s as its components. Become.
The voice amplitude distribution detection unit 14 sets a period during which the voice recognition in-progress signal Ron is on as a calculation execution period. Then, the input speech signal x in the time interval during the calculation execution period other than the speech speech segment SonD represented by the speech speech segment data Son, which is a time segment in which only the noise b is included as a component in the input speech signal x, Speaking using the input speech signal x in the speech speech section SonD represented by the speech speech section data Son during the calculation execution period, which is a time section in which the speech signal x includes the noise b and the speech speech s as components. Calculation of the average amplitude distribution f (n) of the voice s is executed. Here, the details of the method of calculating the average amplitude distribution f (n) of the uttered speech s in the speech amplitude distribution detection unit 14 will be described later.

  Now, the first noise amplitude distribution detection unit 11 and the second noise distribution detection unit repeatedly calculate the amplitude distribution gb (n) using the input speech signal x in a certain unit time interval during the above-described calculation execution period. Whenever the amplitude distribution gb (n) of the noise b is calculated, the content of the noise amplitude distribution register 13 is updated with the calculated amplitude distribution gb (n). When the calculation execution period is less than the above-described unit time interval, the amplitude distribution gb (n) is not calculated during the calculation execution period, and the content of the noise amplitude distribution register 13 is not updated. It will be.

Therefore, the amplitude distribution gb (n) of the noise b stored in the noise amplitude distribution register 13 is always the amplitude distribution gb (bb) of the noise b calculated by the first noise amplitude distribution detector 11 and the second noise distribution detector. n) of the latest amplitude b of the latest noise b calculated.
Next, the average amplitude distribution f (n) of the utterance voice s in the voice amplitude distribution detection unit 14 is calculated every time the voice recognition process is executed, and the voice amplitude distribution detection unit 14 calculates the average of the utterance voice s. Each time the amplitude distribution f (n) is calculated, the contents of the audio amplitude distribution register 15 are updated with the calculated amplitude distribution f (n). Therefore, at the start of execution of the speech recognition process, the average amplitude distribution f (n) of the uttered speech s calculated at the previous execution of the speech recognition process is stored in the speech amplitude register.

  Next, the convolution calculator 16 has an amplitude distribution gb (n) of the noise b stored in the noise amplitude distribution register 13 and an average amplitude distribution f (n) of the uttered speech s stored in the voice amplitude register. ) And the amplitude distribution h (n) of the input audio signal from the microphone 4 is calculated. In Equation 1, Smax is the maximum class number of the speech voice s, and Bmax is the maximum class number of the noise b.

Next, the input gain control unit 10 controls the input gain G (dB) of the input amplifier 5 by the input gain control process shown in FIG.
Assume that the range of input speech data that can be properly processed by the speech recognition engine 7 is a standard range R, and the standard range is a range from Rmin to Rmax. The dynamic range Rmax / Rmin of the standard range is referred to as the standard value D of the dynamic range of the speech recognition engine 7.

As shown in FIG. 3, in the input gain control process, first, the input gain G of the input amplifier 5 is set to a predetermined minimum gain Gmin (step 302), and during speech recognition output from the speech recognition engine 7. Wait until the signal Ron becomes 1 and the speech recognition process is started (step 304).
Next, when the speech recognition in-process signal Ron is turned on and the speech recognition processing is started (step 304), the amplitude distribution h (n) of the input speech signal from the microphone 4 output from the convolution calculator 16 is reached. Hmax is the maximum value of the amplitude distribution indicated by () (the maximum value of the amplitude where the frequency is present), and Hmin is the minimum value (minimum value of the amplitude where the frequency is present) indicated by the amplitude distribution h (n) (step 306). ), Whether the dynamic range Hmax / Hmin of the input voice signal from the microphone 4 represented by the amplitude distribution h (n) is equal to or less than the standard value D of the dynamic range of the voice recognition engine 7 is checked (step 308).

  If the dynamic range Hmax / Hmin of the input speech signal is equal to or less than the standard value D of the dynamic range of the speech recognition engine 7 (step 308), the center of the range of the input speech signal Hmid = (Hmax + Hmin) / 2 Then, the center Rmid = (Rmax + Rmin) / 2 of the standard range of the input speech data of the speech recognition engine 7 is obtained (step 310).

Next, the input gain G of the input amplifier 5 is set to Rmid / Hmid (step 312).
As a result, when the dynamic range of the standard range of the speech recognition engine 7 is equal to or greater than the dynamic range of the input speech data, the input gain G of the input amplifier 5 is set as follows.
That is, it is assumed that the center Hmid of the amplitude distribution h (n) is at a position shifted from the center Rmid of the standard range R of the speech recognition engine 7, as shown in FIG. Here, the amplitude distribution h (n) is input to the speech recognition engine 7 when an input speech signal having an amplitude distribution equal to the amplitude distribution h (n) is AD converted without amplifying the input speech signal by the input amplifier 5. It matches the amplitude distribution of the input voice data to be input.

  In such a case, according to the setting of the input gain G in step 312, as shown in FIG. 4a2, an input audio signal having an amplitude distribution equal to the amplitude distribution h (n) is input in step 312. Amplitude distribution h in (n) of input voice data amplified by gain G and AD converted, that is, amplitude distribution of input voice data input to voice recognition engine 7 when input gain G is set in step 312 as described above. The center of “hin (n)” coincides with the center Rmid of the standard range R of the speech recognition engine 7. In addition, the input voice signal actually picked up by the microphone 4 during the voice recognition process can be expected to have an amplitude distribution approximate to the amplitude distribution h (n).

  Therefore, according to such setting of the input gain G, the entire amplitude distribution h in (n) of the input speech data input to the speech recognition engine 7 is within the standard range R of the speech recognition engine 7. Fits in the center part. Here, in general, the speech recognition engine 7 can perform speech recognition with high accuracy on input speech data having an amplitude distribution in the central portion within the standard range R of the speech recognition engine 7.

Returning to FIG. 3, if the input gain G is set in step 312, the voice recognition in-progress signal Ron output from the voice recognition engine 7 is turned off and the voice recognition process ends (step 314). Return to the processing from step 302.
On the other hand, if the dynamic range Hmax / Hmin of the force speech signal exceeds the standard value D of the dynamic range of the speech recognition engine 7 (step 308), the dynamic range MD = Mmax / Mmin is the standard range D of the speech recognition engine 7. And the frequency M included in the range on the amplitude distribution h (n) of the input audio signal (the total number of occurrences of the amplitude value included in the range) is maximized. Is calculated (step 316). However, Mmin represents the minimum value of the range M, and Mmax represents the maximum value of the range M.

Next, the input gain G of the input amplifier 5 is set to Rmin / Mmin (step 318).
As a result, when the dynamic range of the standard range of the speech recognition engine 7 is less than the dynamic range of the input speech data, the input gain G of the input amplifier 5 is set as follows.
That is, it is assumed that the amplitude distribution h (n) exists so that its end portion is included in the standard range R of the speech recognition engine 7, as shown in FIG. 4b1. Here, the amplitude distribution h (n) is input to the speech recognition engine 7 when an input speech signal having an amplitude distribution equal to the amplitude distribution h (n) is AD converted without amplifying the input speech signal by the input amplifier 5. It matches the amplitude distribution of the input voice data to be input.

  In such a case, according to the setting of the input gain G in step 318, an input audio signal having an amplitude distribution equal to the amplitude distribution h (n) is input as set in step 318 as shown in FIG. 4b2. Amplitude distribution h in (n) of input voice data amplified by gain G and AD-converted, that is, amplitude distribution of input voice data input to voice recognition engine 7 when input gain G is set in step 318 as described above. Hin (n) has the maximum frequency (appearance probability) of the amplitude value within the standard range R of the speech recognition engine 7. In addition, the input voice signal actually picked up by the microphone 4 during the voice recognition process can be expected to have an amplitude distribution approximate to the amplitude distribution h (n).

  Therefore, according to such setting of the input gain G, the input speech data input to the speech recognition engine 7 has a range of amplitude values with a high frequency (appearance probability), that is, a range of amplitude values considered to be main. The voice recognition engine 7 is within the standard range R, so that the voice recognition engine 7 can perform voice recognition satisfactorily.

Returning to FIG. 3, if the input gain G is set in step 318, the process waits until the speech recognition processing signal Ron output from the speech recognition engine 7 is turned off and the speech recognition process ends (step 314). ), The process returns to step 302.
The input gain control process has been described above.
Note that the reason why the input gain G is set to the minimum gain Gmin in the above step 302 is that the input voice signal is saturated by the amplification of the input amplifier 5 during the period when the voice recognition process is not performed. This is to prevent the noise amplitude distribution gb (n) from being appropriately calculated in the first noise amplitude distribution detecting unit 11 that calculates the noise amplitude distribution gb (n) during the period. The minimum gain Gmin is, for example, a value that allows the AD converter 6 to convert a maximum volume of sound that can be picked up without distortion by the microphone 4 to a maximum value that can be expressed as input sound data. To do.

Next, a calculation method of the average amplitude distribution f (n) of the speech s of the speech amplitude distribution detection unit 14 described above, a calculation method of the amplitude distribution gb (n) of the noise b of the first noise amplitude distribution detection unit 11, A method for calculating the amplitude distribution gb (n) of the noise b of the second noise amplitude distribution detector 12 will be described.
First, a method for calculating the average amplitude distribution f (n) of the uttered voice s by the voice amplitude distribution detector 14 will be described.
It is known that the amplitude distribution f (s) of the uttered speech s is a super Gaussian distribution, and assuming that the amplitude distribution of the uttered speech s is a super Gaussian distribution, The amplitude distribution can be represented.

  Here, α and β in Equation 2 have the relationship of Equation 3 with the average μs and variance σs of the speech s.

  Further, the average μs and the variance σs have the relationship of the power (root mean square) Ps of the uttered voice s and Equation 4.

  Therefore, the relationship between α and β in Expression 2 and the power Ps of the speech voice s can be expressed by Expression 5.

Here, the peak of the amplitude distribution of the uttered voice s appears near 0, and in this case, α can be set to approximately 1 as shown in the following references.
References: T. Lotter and P. Vary, “Noise reduction by joint maximum a posteriori spectral amplitude and phase estimation with super-gaussian speech modeling”, Proc. EUSIPCO-04 (Vienna, Austria), pp. 1447-60, Sep. .2004.
If α = 1, the relationship between β and power Ps can be expressed by Equation 6, and if β is obtained, the amplitude distribution f (s) of the uttered speech s in Equation 1 can be calculated.

Therefore, the voice amplitude distribution detection unit 14 obtains and stores the power Pse of the uttered voice s for each calculation execution period. Here, the calculation of the power Pse of the speech s during the calculation execution period is performed as follows.
That is, the input audio data gain-adjusted by a gain / G that can be expressed by the reciprocal of the input gain G of the input amplifier 5 used when generating the input audio data is used as the target input. Let it be audio data. Here, the target input audio data represents the value of the input audio signal x before being amplified by the input amplifier 5.

Then, the voice amplitude distribution detection unit 14 obtains and stores the target input voice data during the calculation execution period, and uses the saved target input voice data to calculate the power Pse of the uttered voice s as follows. Calculate and save.
That is, during the calculation execution period, the period other than the speech voice section represented by the speech voice section data Son is a time section in which only the noise b is included as a component in the input voice signal x. The power of the input audio data is calculated as power Pb. Further, the power of the target input voice data in the utterance voice section in which noise b and utterance voice s are included as components in the input voice signal x is calculated as power Pb + s. Then, the power Pb of the uttered voice s is calculated and stored by subtracting the power Pb from the power Pb + s.

  Then, when calculation and storage of the power Pse of each speech s is completed, β is obtained as the power Ps of Equation 6 by using the average power Pse of the speech s stored so far, and the speech is obtained from the obtained β. The amplitude distribution f (s) of the voice s is calculated. Then, the amplitude distribution f (s) is discretized to obtain an average amplitude distribution f (n) of the speech voice s.

Next, a method for calculating the amplitude distribution gb (n) of the noise b of the first noise amplitude distribution detector 11 will be described.
First, during the calculation execution period of the first noise amplitude distribution detection unit 11, the speaker output sound a and the noise b are included as components in the input sound signal x.
Therefore, the amplitude distribution gc (n) of the input audio signal x is expressed by a convolution operation represented by Equations 7 and 8 between the amplitude distribution ga (n) of the speaker output sound a and the amplitude distribution gb (n) of the noise b. Can do.

In Expression 8, Amax is the number of the maximum value of the speaker output sound a, and Bmax is the number of the maximum value of the noise b.
Then, as shown in Equation 9, W representing the amplitude distribution gb (n) of the noise b and the amplitude distribution ga (n) of the speaker output sound a representing the matrix are determined.

  In this case, it can be seen from Equations 7 and 8 that W that minimizes the mean square error J in the unit time interval of error e shown in Equation 10 is the true value of W. Note that E [X] represents an average value of the unit time interval of X.

  Then, W that minimizes the mean square error J is obtained as shown in Equation 11, where W is a value obtained by partial differentiation of the mean square error J with W.

The amplitude distribution gb (n) of the noise b is determined from W that minimizes the mean square error J.
Accordingly, the first noise amplitude distribution detection unit 11 can represent the gain of the input voice data AD-converted by the AD converter 6 by a gain / G that can be expressed by the reciprocal of the input gain G of the input amplifier 5 used when the input voice data is generated. The first noise amplitude distribution detecting unit 11 obtains and stores the target input voice data during the calculation execution period, and the audio data to be input to the DA converter 1 as the target input voice data. save.
Then, during the calculation execution period, the amplitude distribution gb (n) of the noise b is calculated for each unit time section as follows using the stored target input voice data and audio data.
That is, the transfer function H from the input of the DA converter 1 to the output of the microphone 4 is calculated with reference to the output gain SpG of the output amplifier 2, and the transfer function H calculated for the audio data input to the DA converter 1 is calculated. The amplitude distribution function of the unit time interval of the applied audio data is calculated as the amplitude distribution ga (n) of the speaker output sound a. For example, the transfer function H is obtained by adding the output gain SpG of the output amplifier 2 to the transfer function obtained in advance from the input of the DA converter 1 to the output of the microphone 4 when amplification is not performed by the output amplifier 2. Find by multiplying. Alternatively, the transfer function H can be obtained in real time from target input voice data and audio data using an adaptive filter or the like.

  Further, the amplitude distribution of the target input voice data in the unit time interval is calculated as the amplitude distribution gc (n) of the input voice signal x.

Then, the amplitude distribution gb (n) of the noise b is calculated from the amplitude distribution ga (n) and the amplitude distribution gc (n) calculated as described above during the unit time according to Equation 11.
Next, a method for calculating the amplitude distribution gb (n) of the noise b of the second noise amplitude distribution detector 12 will be described.
Input audio data whose gain is adjusted by a gain / G that represents the gain of the input audio data AD-converted by the AD converter 6 by the reciprocal of the input gain G of the input amplifier 5 used when generating the input audio data. As described below, the second noise amplitude distribution detector 12 obtains and stores the target input voice data during the calculation execution period, and uses the saved target input voice data for each unit time interval as follows. Then, the amplitude distribution gb (n) of the noise b is calculated.
That is, during the calculation execution period of the second noise amplitude distribution detector 12, the input audio signal x includes only the noise b as a component. Therefore, the second noise amplitude distribution detection unit 12 calculates the amplitude distribution of the target input speech data in the unit time interval as the amplitude distribution gb (n) of the noise b as it is.

The embodiment of the present invention has been described above.
As described above, according to the present embodiment, the first noise amplitude distribution detection unit 11 and the second amplitude distribution detection unit repeatedly execute the calculation of the noise amplitude distribution gb (n) in a time section other than the speech voice section. At the start of the speech recognition process, the noise amplitude distribution gb (n) detected last, that is, the noise amplitude distribution gb (n) at the most recent time point, and the utterance speech at the time of the previous speech recognition process execution The amplitude distribution h (n) of the input speech signal is estimated based on the average amplitude distribution f (n) of the input amplitude, and the amplitude obtained by amplifying the estimated amplitude distribution h (n) of the input speech signal by the input amplifier 5 The input gain G of the input amplifier 5 is set so that the distribution h in (n) becomes a level suitable for the voice recognition engine 7.

  The noise amplitude distribution gb (n) at the most recent time can be expected to approximate the noise amplitude distribution gb (n) in the current noise situation. Therefore, according to such a speech recognition apparatus, an input gain G more suitable for the current noise situation can be set in the input amplifier 5 at the start of speech recognition processing.

  By the way, in the above embodiment, the input gain G of the input amplifier 5 is set based on the amplitude distribution, but this is another characteristic value Z (Z is an example of the voice level instead of the amplitude distribution). The input gain G of the input amplifier 5 can be set based on the distribution of the peak value of amplitude, the maximum amplitude value, the average amplitude, or the like.

  That is, in this case, the first noise amplitude distribution detection unit 11 and the second amplitude distribution detection unit calculate the noise characteristic value Z instead of the noise amplitude distribution gb (n), and the voice amplitude distribution detection unit. 14, the average characteristic value Z of the uttered voice is calculated, and the input voice signal is calculated based on the noise characteristic value Z finally calculated by the input gain control unit 10 and the average characteristic value Z of the uttered voice. The characteristic value Z of the input amplifier 5 is estimated based on the estimated characteristic Z of the input voice signal, and the input gain G of the input amplifier 5 is determined according to a rule predetermined according to the type of characteristic value to be the characteristic value Z. Is set so that the amplitude range of the input voice data obtained by amplifying the signal by the input gain G and AD-converting it matches the standard range R of the voice recognition engine 7.

  DESCRIPTION OF SYMBOLS 1 ... DA converter, 2 ... Output amplifier, 3 ... Speaker, 4 ... Microphone, 5 ... Input amplifier, 6 ... AD converter, 7 ... Speech recognition engine, 8 ... Talk switch, 9 ... Output gain control part, 10 ... Input gain control unit, 11 ... first noise amplitude distribution detection unit, 12 ... second noise amplitude distribution detection unit, 13 ... noise amplitude distribution register, 14 ... audio amplitude distribution detection unit, 15 ... audio amplitude distribution register, 16 ... tatami Calculation unit, 17... Gain control unit.

Claims (6)

A speech recognition device that performs speech recognition,
With a microphone,
An input amplifier that amplifies the input audio signal output from the microphone;
An AD converter that converts the signal amplified by the input amplifier into input audio data;
A speech recognition engine that performs speech recognition processing on input speech data output from the AD converter in response to a speech recognition execution instruction;
A noise level detector;
An utterance voice level detector;
An input gain control unit for controlling the gain of the input amplifier;
In the voice recognition process, the voice recognition engine detects a time interval in which the input voice data includes a user's utterance voice as a utterance voice section, and the utterance voice included in the input voice data in the detected utterance voice section. Identify the content,
The noise level detector repeats a noise level included in the input voice signal based on the input voice data in a time section other than the speech voice section or a time section in which the voice recognition process is not performed. Calculate
The utterance voice level detection unit calculates an average level of the utterance voice included in the input voice signal of each utterance voice section detected in each time of the voice recognition processing based on the input voice data,
The input gain control unit, at the start of each time of the speech recognition processing, the noise level calculated last by the noise level detection unit, and the average level of the utterance speech detected by the utterance speech level detection unit Thus, the level of the input voice signal in the utterance voice section detected in the voice recognition process of the time is estimated, and the level obtained by amplifying the estimated level of the input voice signal with the input amplifier is the voice recognition engine. A speech recognition apparatus, wherein a gain of the input amplifier is set so as to be a level suitable for. The speech recognition apparatus according to claim 1,
The input gain control unit is configured such that, during a time period when the speech recognition process is not performed, the gain of the input amplifier is amplified by the input amplifier so that the maximum level that can be taken by the level of the input speech signal is the AD amplifier. A speech recognition apparatus, wherein a predetermined value is set so as not to exceed an input range of the converter. The speech recognition apparatus according to claim 1 or 2,
An audio device that outputs the audio sound represented by the audio data;
An output suppression unit that suppresses output of audio sound of the audio device during a period in which the speech recognition process is performed;
The noise level detection unit calculates a noise level included in the input voice signal based on the input voice data and the audio data in a time interval in which the voice recognition process is not performed. Voice recognition device. The speech recognition device according to claim 1, 2, or 3,
The noise level detector calculates a noise amplitude distribution as the noise level,
The utterance voice level detection unit calculates an average amplitude distribution of the utterance voice as an average level of the utterance voice,
The speech recognition apparatus, wherein the input gain control unit estimates an amplitude distribution of the input speech signal as a level of the input speech signal. The speech recognition device according to claim 4,
When the dynamic range of the amplitude distribution range indicated by the amplitude distribution of the estimated input speech signal is equal to or less than the dynamic range of the input range of the speech recognition engine, the input gain control unit may determine the estimated input speech signal. The gain of the input amplifier is adjusted so that the amplitude value after the amplitude value at the center of the amplitude distribution range indicated by the amplitude distribution is amplified by the input amplifier becomes the amplitude value at the center of the input range of the speech recognition engine. A speech recognition apparatus characterized by setting. The speech recognition device according to claim 4 or 5,
When the dynamic range of the amplitude distribution range in the estimated amplitude distribution of the input speech signal exceeds the dynamic range of the input range of the speech recognition engine, the input gain control unit determines the amplitude of the estimated input speech signal. of distribution range of the amplitude in the distribution, in the range moiety having the same dynamic range as the dynamic range of the input range of the voice recognition engine, it selects a range that the total amount of power within that range portion becomes maximum, A speech recognition apparatus, wherein a gain of the input amplifier is set so that a range after the selected range portion is amplified by the input amplifier matches an input range of the speech recognition engine.