JPH02293797A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To lessen erroneous recognition even in low S/N environment by calculating the S/N which is the ratio of an input voice power and a noise power, determining the prescribed threshold according to the high and low values of the S/N and executing the section detection of the input voice and the section correction of registered voices, then executing collating. CONSTITUTION:The background noise power of every unit time is calculated in an analyzing section 2 and the threshold is set in a temporarily threshold setting section 3 by adding the prescribed value to the average value thereof. The power of the input voice is thereafter calculated and the S/N is calculated in an S/N calculating section 4. The threshold setting section 5 sets the value obtd. after the prescribed value is subtracted from the peak value as the threshold when the S/N is above the prescribed value. The threshold of the setting section 3 is set as the threshold for the region where the S/N is low. The section of the contents in a buffer 10 for input voice is detected in a section detecting section 6 by this threshold. The section correction is not executed and the collating of the contents of the buffer 9 and buffer 10 for registered voice is executed without executing the section correction when the S/N is large. The contents of the buffer 9 are corrected in the section correcting section and the collating is executed when the S/N is small. The deviation in the collating is eliminated in this way and the erroneous recognition is lessened even in the low S/N environment.

Description

[Detailed description of the invention] Industrial applications The present invention relates to a speech recognition device.

Background of the Invention In recent years, with the development of speech recognition technology, speech recognition devices are being put into practical use in various fields. It needs to be resolved. One of the practical problems is that when the SN ratio at the time of voice input is low, noise is erroneously detected as a voice section, resulting in erroneous recognition.

In conventional speech recognition devices, in order to solve the above problem, background noise is input in advance just before inputting the speech, its power is checked, and the threshold value for detecting speech sections is set above the above power. Then, the voice section is detected using the set threshold.

With this method, even in an environment with a low SN ratio, noise will not be erroneously detected as a speech section, and the erroneous recognition rate will be reduced.

Hereinafter, a conventional speech recognition device as described above will be explained with reference to the drawings. FIG. 3 is a block diagram of a conventional registered word speech recognition device. In the same figure,
1 is an audio input terminal, 2 is an analysis section, 16 is a threshold setting section, 1
7 is a section detection section, 18 is a collation section, 19 is a recognition result output terminal, 20 is a buffer for registered speech, 21 is a buffer for input speech, and 22 and 23 are switches. The operation of the speech recognition device configured as described above will be explained below.

First, at the time of registration, a background noise signal for a predetermined period of time is input from the audio input terminal 1 immediately before audio input, the power of the signal for each unit time is calculated by the analysis section 2, and the calculation result is input to the threshold value setting section 16. be done. The threshold value setting unit 16 calculates the average value of the powers calculated above, and adds a predetermined value (
(for example, 6 dB) is set as the section detection threshold.

When inputting registered word audio, the analysis unit 2 calculates the power and characteristic parameters of the signal for each unit time based on the signal input from the audio input terminal 1.The power calculation result is sent to the section detection unit 17, and the characteristic parameters are input. It is input to the audio buffer 21. If, for example, the LPC cepstral method is used as the analysis method, a predetermined number of cepstral coefficients are calculated as characteristic parameters. Next, the section detection section 17
Then, the signal power for each unit time is compared with the section detection threshold set previously, and a portion where the signal power is continuously equal to or higher than the section detection threshold for 60 msec or more is determined to be a voice section. However, even if the signal power becomes below the section detection threshold, if the section where the signal power is below the threshold does not continue for 60 msec or more, it is determined as a voice section. Next, the characteristic parameters for the determined voice section are inputted from the input voice buffer 21 and stored in the registered voice buffer 20. The process after the voice input described above is repeated for all recognized words.

Next, during recognition, after setting a section detection threshold based on background noise as in the case of registration, the input speech is analyzed and speech sections are detected. Both the analysis method and section detection method are the same as at the time of registration. After the voice section is detected, the collation unit 18 collates the registered voice and the input voice, and outputs the word indicating the shortest distance from the recognition result output terminal 19 as a recognition result. Note that the switch 22 is set to the threshold value setting section 1 when inputting noise immediately before voice input.
6, when inputting audio, the calculation result is input to the section detecting section 17 and the input audio buffer 21. The switch 23 operates to input the characteristic parameters to the registration buffer 20 during registration and to the matching unit 18 during recognition.

Problem to be Solved by the Invention However, with the above configuration, it is possible to remove noise regardless of changes in noise power, but the start and end positions of the voice section may change due to changes in noise or vocalization power. Because of this shift, if the conditions are different between when the registered voice or standard voice is uttered and when the input voice is uttered, different voice sections are compared, resulting in a problem that misrecognition is likely to occur.

In view of the above-mentioned problems, the present invention provides a speech recognition device that can prevent deviations in speech intervals between registered speech or standard speech and input speech, and can reduce misrecognition caused by differences in situations.

Means for Solving the Problems In order to achieve the above object, the speech recognition device according to claim 1 includes an analysis section that detects the power of the input signal per unit time, and an SN ratio calculation section that calculates the SN ratio. a threshold setting section that determines a section detection threshold in consideration of the SN ratio; a section detection section that detects a speech section of the input signal using the determined threshold; and a section correction section that corrects the registered speech or standard speech section. and a matching unit that matches the registered speech or standard speech with the input speech and outputs a recognition result.

The speech recognition device according to claim 2 further includes an analysis section that detects the power of the input signal per unit time, and a threshold setting section that sets the section detection threshold in consideration of the peak value of the signal power and the noise power value. , a section detection section that detects the speech section of the input signal using a set threshold; a section correction section that corrects the section of the registered speech or standard speech; and a recognition result by comparing the registered speech or standard speech with the input speech. It consists of a collation section that outputs.

According to the speech recognition device according to claim 1, the SN ratio calculation section calculates the SN ratio, and the threshold value determination section sets a value higher than the noise power in an environment where the SN ratio is low, and sets a value equal to or higher than the noise power in an environment where the SN ratio is high. After determining a value obtained by subtracting a predetermined value from the value as a threshold, a section detecting section detects sections of the input speech using the above threshold, and a section correcting section corrects the registered speech or standard speech section using the above threshold. , the collation unit collates the registered voice or standard voice with the input voice.

According to the speech recognition device according to claim 2, the threshold value setting section compares the value obtained by adding the predetermined value to the noise power value and the value obtained by subtracting the predetermined value from the peak value, and sets the larger value as the threshold value. After setting, the section detecting section detects sections of the input speech using the above thresholds, the section correcting section corrects the registered speech or standard speech section using the above thresholds, and the matching section compares the sections with the registered speech or standard speech. Verify with input audio.

Embodiment FIG. 1 is a block diagram of a registered word speech recognition device in a first embodiment of the present invention (corresponding to the invention set forth in claim 1).

In the figure, 1 is an audio input terminal, 2 is an analysis section, 3 is a temporary threshold setting section, 4 is an SN ratio calculation section, 5 is a threshold setting section, 6 is an interval detection section, 7 is an interval correction section, and 8 is a verification section. 9 is a buffer for registered audio, 10 is a buffer for input audio, 11 is a recognition result output terminal, 12 and 13 are switches, and the conventional example (
(See Figure 3) are given the same numbers.

The operation of the speech recognition device configured as described above will be explained below.

First, at the time of registration, a background noise signal for a predetermined period of time is input from the audio input terminal 1 immediately before audio input, and the power of the signal for each unit time is calculated by the analysis section 2. The calculation result is input to the temporary threshold setting section 3.

The temporary threshold setting unit 3 calculates the average value of the power, and sets the value obtained by adding a predetermined value (8 dB in this embodiment) to the average value as the temporary section detection threshold.

At the time of voice input of registered words, the analysis section 2 calculates the power and characteristic parameters of the signal for each unit time based on the signal input from the voice input terminal 1. The power calculation result is input to the SN ratio calculation unit 4, and the feature parameters are input to the input audio buffer 10.

Note that the analysis method is the same as in the conventional example. In the SN ratio calculation unit 4, the signal portion that is equal to or higher than the previously set temporary section detection threshold is regarded as a temporary voice section, and the ratio of the average value of the peak value within the temporary voice section and the previously calculated noise power is determined as the SN ratio. Calculated as a ratio,
If the SN ratio is less than a predetermined value (24 dB in this example), the speaker is instructed to perform the registration again, and the above registration process is restarted from the beginning.

If the SN value is 24 dB or more, the threshold setting unit 5 determines a value obtained by subtracting a predetermined value (18 dB in this embodiment) from the peak value as the detection threshold. A section detection unit 6 compares the signal power for each unit time with a detection threshold to detect a voice section. The section detection method is the same as in the conventional example. Next, the characteristic parameters for the voice section are inputted from the input voice buffer 10 and registered in the registered voice buffer 9. The process after the above registered voice input is repeated for all recognized words.

Next, at the time of recognition, after setting a temporary section detection threshold based on background noise as in the case of registration, the input voice is analyzed and the results are input to the SN ratio calculation unit 4 and the input voice buffer 10. SN
The ratio calculation section 4 calculates the SN ratio in the same manner as during registration, and inputs the result into the threshold setting section 5.

In the threshold value setting section 5, if the SN ratio is 24 dB or more, the value obtained by subtracting 18 dB from the peak value is set as the threshold value, and the SN ratio is set to 24 dB.
If it is less than or equal to B, the previous provisional section detection threshold is set as the threshold, and then the section detecting section 6 detects the voice section using the threshold described on the left. Note that the section detection method is the same as at the time of registration.

Next, the section correction unit 7 does not correct the registered voice section when the SN ratio is 24 dB or more, and only when the SN ratio is 24 dB or less, detects the section of the registered voice again using the threshold value. Start over. Next, the verification section 8 performs verification against the registered voices of seven people, and outputs the word indicating the shortest distance from the output terminal 11 as a recognition result. Note that the switch 12 is
It operates to input the calculation result to the temporary threshold setting section 3 when inputting noise immediately before inputting the voice, and to input the calculation result to the SN ratio calculation section 4 and the incoming voice buffer 10 when inputting the voice.

The switch 13 operates to input the feature parameters into the registration buffer 9 during registration and into the section correction unit 7 during recognition.

As described above, according to this embodiment, the SN ratio calculation section 4 calculates the ratio between the peak value of the signal and the average power value of the noise, and the threshold setting section 5 calculates the ratio when the SN ratio is below a certain value. A value obtained by adding a predetermined value to the noise power value is determined as a threshold value, and a value obtained by subtracting a predetermined value from the peak value when the S/N ratio is greater than a predetermined value is determined as a threshold value. By performing section detection, correcting the registered speech using the above-mentioned threshold value in the section correction section 7, and comparing the above-mentioned registered speech and input speech in the matching section 8, the deviation in the speech section between the registered speech and the input speech is corrected. prevent,
Misrecognition due to differences in situations can be reduced.

FIG. 2 is a block diagram of a registered word speech recognition device according to a second embodiment of the present invention (corresponding to the invention set forth in claim 2).

In the figure, 1 is an audio input terminal, 2 is an analysis unit, 3 is a temporary threshold setting unit, 14 is a threshold setting unit, 6 is an interval detection unit, 7 is an interval correction unit, 8 is a matching unit, and 9 is for registered audio. Bachha 1
0 is input audio buffer, 11 is recognition result output terminal, 1
2 and 15 are switches, which are the same as those in the previous embodiment.
The same number is assigned.

The operation of the speech recognition device configured as described above will be explained below.

First, at the time of registration, a background noise signal for a predetermined period of time is input from the audio input terminal 1 immediately before audio input, and the power of the signal for each unit time is calculated by the analysis section 2. The calculation result is input to the temporary threshold setting section 3.

The temporary threshold setting unit 3 calculates the average value of the power, and sets the value obtained by adding a constant value (6 dB in this embodiment) to the average value as the temporary section detection threshold.

When registering word speech is input, the analysis section 2 calculates the signal power and characteristic parameters for each unit time based on the signal input from the speech input terminal 1. The power calculation result is input to the threshold value setting section 14, and the feature parameters are input to the input audio buffer 10. Note that the analysis method was the same as in the previous example. The threshold setting unit 14 defines a signal portion that is equal to or higher than the previously set temporary section detection threshold as a temporary voice section, and calculates a predetermined value (18 dB in this embodiment) from the peak value within the temporary voice section.
is compared with the provisional section detection threshold calculated earlier, and if the latter value is larger, the speaker is instructed to redo the registration, and the above registration process is restarted from the beginning.

If the former value is large, this former value (peak value - 18
dB) is set as a detection threshold, and the section detection unit 6 compares the signal power for each unit time with the detection threshold to detect a voice section. The section detection method is the same as in the conventional example. next,
Audio buffer 10 for inputting feature parameters for audio sections
and register it in the registered audio buffer 9. The above process after the registered voice is repeated for all recognized words.

Next, during recognition, after setting a temporary section detection threshold based on background noise in the same way as during registration, the input speech is analyzed, and the results are input to the threshold setting unit 14 and the input speech buffer 10. The threshold setting unit 14 defines the signal portion that is equal to or higher than the section detection threshold as a provisional speech section, and compares the value obtained by subtracting 18 dB from the peak value within the provisional speech section with the provisional section detection threshold calculated previously, and calculates the difference between the two values. After setting the larger one as a threshold, the section detecting section 6 detects a speech section using the threshold described on the left. Note that the section detection method is the same as at the time of registration.

Next, the section correction unit 7 sets the threshold value to 18 d from the peak value.
If the value is set by subtracting B, the registered voice section is not corrected, and only when the threshold is set at the temporary section detection threshold, the section detection of the registered voice is re-performed using the threshold. Next, the matching section 8 matches the registered speech and the input speech, and outputs the word indicating the shortest distance from the output terminal 11 as a recognition result. Note that the switch 15 operates to input the calculation result to the temporary threshold setting section 3 when inputting noise immediately before voice input, and to input the calculation result to the threshold setting section 14 and the input voice buffer 10 when inputting voice. The switch 13 operates to input the feature parameters to the registration buffer 9 during registration and to the section correction unit 7 during recognition.

As described above, according to the present embodiment, the threshold value setting unit 14 uses a value obtained by adding 6 dB to the noise power value and a value of 18 dB from the peak value.
After comparing the value obtained by subtracting B and determining the larger l5 as the threshold, the section detecting section 6 detects sections of the input speech using the above threshold, and the section correcting section detects the registered speech section using the above threshold. By correcting the registered speech and comparing the input speech with the registered speech in the matching section, it is possible to prevent a shift in the speech section between the registered speech and the input speech, and to reduce misrecognition due to differences in situations. Further, in this embodiment, compared to the first embodiment, the same effect can be expected without taking the trouble of calculating the SN ratio.

Effects of the Invention In the speech recognition device according to claim 1, the SN ratio calculation section calculates the SN ratio, and the threshold value determination section sets a value higher than the noise power in an environment with a low SN ratio, and a value higher than the peak value in an environment with a high SN ratio. After setting a value obtained by subtracting a predetermined value as a threshold, the section detecting section detects sections of the input speech using the above threshold, and the section correcting section corrects the registered speech or standard speech section using the above threshold and performs verification. By comparing the registered speech or standard speech with the input speech in the section, it is possible to prevent a shift in the speech interval between the registered speech and the input speech, and to reduce misrecognitions 1[i-] due to differences in situations.

Further, in the speech recognition device according to claim 2, after the threshold value setting section compares the value obtained by adding the predetermined value to the noise power value and the value obtained by subtracting the predetermined value from the peak value, and determines the larger one as the threshold value, The section detecting section detects sections of the input speech using the above threshold, the section correcting section corrects the registered speech or standard speech section using the above threshold, and the matching section compares the registered speech or standard speech with the input speech. By performing the comparison, it is possible to prevent a shift in the voice section between the registered voice or standard voice and the input voice, and to reduce misrecognition due to differences in situations. Moreover, compared to the above invention, the same effect can be expected without taking the trouble of calculating the SN ratio.

[Brief Description of the Drawings] Fig. 1 is a block diagram of a speech recognition device in a first embodiment of the present invention, Fig. 2 is a block diagram of a speech recognition device in a second embodiment of the invention, and Fig. 3 1 is a block diagram of a conventional speech recognition device. ．． Threshold value setting unit, e. ．． ．． Section detection unit, 7. ．． ．． Section correction section, 8. ．． ．． Collation section. Name of agent: Patent attorney Shigetaka Awano 1 person 2. ．． ．． Analysis Department, 4. ．． ．． S/N ratio calculation unit, 5, 14. ．． Figure I

Claims (2)

[Claims] (1) An analysis unit that detects the power of the input signal per unit time and the ratio of voice power to noise power (hereinafter referred to as SN ratio)
a threshold determination unit that determines a section detection threshold in consideration of the SN ratio; a section detection section that detects a voice section of the input signal using the determined threshold; and a registered voice or standard voice. The section includes a section correction section that corrects the section of the input speech, and a matching section that matches the registered speech or standard speech with the input speech and outputs a recognition result.The SN ratio calculation section calculates the SN ratio, and the threshold value determination section In an environment with a low S/N ratio, set the value above the noise power, and in an environment with a high S/N ratio, set the maximum power value of the signal (
A threshold value is set as a value obtained by subtracting a predetermined value from the peak value (hereinafter referred to as the peak value), a section detecting section detects sections of input speech using the above threshold, and a section correcting section uses the above thresholds to detect registered speech or standard speech sections. A speech recognition device characterized in that the speech recognition device is configured such that the matching section matches the registered speech or standard speech with the input speech. (2) An analysis section that detects the power of the input signal per unit time, a threshold setting section that sets an interval detection threshold considering the peak value of the signal power and the noise power value, and the input signal according to the set threshold. Equipped with a section detection section that detects a speech section of a signal, a section correction section that corrects a section of registered speech or standard speech, and a matching section that matches input speech with registered speech or standard speech and outputs a recognition result. Then, the threshold setting section compares the value obtained by adding a predetermined value to the noise power and the value obtained by subtracting the predetermined value from the peak value of the signal, and sets the larger value as the threshold, and the section detection section sets the above threshold value. The section is configured such that the section of the input speech is detected using the above-described method, the section correcting section corrects the registered speech or standard speech section using the threshold value, and the matching section matches the registered speech or standard speech with the input speech. A voice recognition device featuring: