JPS59123896A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

Detailed Description of the Invention (A) Technical Field of the Invention The present invention relates to a speech recognition method, particularly a speech recognition method that performs frequency analysis of input speech using a group of bandpass filters and recognizes speech such as monosyllables or words. .

The present invention relates to a voice recognition method that reduces the amount of time-series feature parameter parameters to be compared and reduces the amount of analysis hardware without reducing the voice recognition rate.

(Bl Technology Background and Problems As a speech recognition method, in order to perform wideband speech frequency analysis, a multi-channel bandpass filter is used, and the output of each filter is converted into band-specific spectral power by rectification and integration, etc.). For spectrum normalization 1, the logarithmic spectral power for each band is converted so that the average value of all channels is zero, and then the logarithmic spectral power for each band is calculated by logarithmically transforming The logarithmic spectral power is used as the feature parameter time series for matching.

A method is known in which speech recognition, such as a single syllable or a word, is performed by comparing a time series of standard feature parameters registered in advance in a dictionary using, for example, a dynamic programmatic (DP) matching method.

In the above speech recognition method, in order to increase the speech recognition rate, it is necessary to increase the number of bandpass filters, that is, the number of channels. However, increasing the number of channels not only requires a large amount of hardware to analyze audio frequencies, but also increases the number of feature parameter elements, which requires a large amount of memory for matching. A large amount of storage space is also required for the standard feature parameter time series stored in the . Furthermore, it takes a lot of time to process the computation for verification.

However, if the number of channels is reduced, the amount of important memory, etc. can be reduced, but the speech recognition rate will deteriorate.

By the way, 1) The inventors of the present invention have conducted many experiments and researches prior to completing the present invention, and as a result, the following results regarding voice recognition have been discovered. I discovered temporality. Better results can be obtained when performing audio frequency analysis over a wide range of frequencies, including the high frequency range.

Especially in the high frequency band part, the relative amount of audio energy in the power spectrum for each sampling is important.9 For example, the peak of the power spectrum is t5.
It is in the frequency part of Klz.

7 K) Is the voice RN6 in the frequency part of Iz? That said, it's not that important. This is probably because it is difficult for two human ears to recognize slight differences in frequencies in high frequency bands.

Therefore, the present inventors normalized the parameters analyzed using multiple band filters, including the high frequency band, and then removed the parameters of multiple channels in the high frequency band, and investigated whether the speech recognition rate would change. In an experiment, it was found that the recognition rate did not decrease compared to the case where parameters for all channels including the high frequency portion were used as the 9% parameter time series. On the other hand, if the high frequency band part is not added to the normalization conditions from the beginning.

It was found that the speech recognition rate decreased.

(q) Purpose and Structure of the Invention In view of the above points, the present invention aims to improve the conventional method, and reduces the amount of feature parameters to be compared to reduce the amount of memory, etc., without reducing the speech recognition rate. The aim is to reduce the amount of hardware required for spectrum analysis.In other words, the aim is to further improve the speech recognition rate with the same amount of feature parameters as before. Therefore, in the speech recognition method of the present invention, which recognizes speech by collating feature parameter time series obtained by frequency analysis of speech, spectrum analysis is performed on a wide speech frequency band. The high frequency band portion of the analysis result is weighted to obtain a normalized spectrum, and then the spectrum from which the high frequency band portion is removed is used as the feature parameter time series for matching.The following 2
An embodiment will be described with reference to the drawings.

(D) Examples of the invention The figure shows an embodiment of the invention.

In the figure, 1 is a voice input section, and 2 is a parameter extraction section.

3 is a spectrum analyzer, 4-1 to 4-27 are band pass filters, and 5-1 to 5-71 are rectifiers.

6-1 to 6-n are analog-to-digital conversion circuits;
7 is a spectrum normalization unit, 8-1 to 8-n are logarithmic conversion units, 9 is a constant storage unit, io is a multiplier, 11 is an average value calculation unit, and 12-1 to 12-(yt-1) are subtraction units. vessel, 1
3 represents a voice recognition section, and 14 represents a speech recognition section 4.

A voice analog signal consisting of a single syllable or word inputted from the voice input section 1 is inputted to the parameter extraction section 2 . The parameter extraction unit 2 performs frequency analysis of the audio analog signal and extracts and generates a time series of characteristic parameters of the input audio to be recognized. Therefore, the parameter extraction unit 2 normalizes the output of the spectrum analysis unit 3 that performs spectrum analysis over a wide audio frequency band and the output of the spectrum analysis unit 3 with respect to the high frequency band. The feature parameter for matching the normalized spectrum excluding the part Fly "2+ ...P
? )-1.

The spectrum analysis section 3 has a plurality of (yz) bandpass filters 4-1 to 4-n for each band.

In the figure, the 2-pass frequency increases in order from the bandpass filter 4-1 at the top to the bottom. The band pass filters 4-1 to 4-n are arranged such that the 3 dB attenuation points of the band pass filters that are in contact with each other coincide with each other, and cover a wide band from, for example, 180 Hz to 7.8 KHz. In particular, the bandpass filters 4-1 to 4-(n-1) have a bandwidth of, for example, 170Hz to 620Hz, but the bandpass filter 4-n in the highest frequency band has a bandwidth of, for example, 3K.
It is designed to have wide band characteristics such as 11z.

The audio signals from audio input units 1 to 9 are passed through a bandpass filter 4-
1 to 4-n for each band.

The signals are input to rectifiers 5-1 to 5-71, respectively. Each rectifier 5-1 to 5-n performs rectification and smoothing of nine human input signals with a rectification and integration time constant of 2, for example, 10772.9. The outputs of the rectifiers 5-1 to 5-21 are input to analog-to-digital converters 6-1 to 6-22, and the spectrum power for each band is obtained as a digital quantity. The conversion result is output to the spectrum normalization section 7.

The band-specific spectral power input to the spectrum normalization unit 7 is logarithmically converted by the logarithmic conversion units 8-1 to B-n so that the output value is proportional to the intensity of sound felt by two people, and is converted into band-specific logarithmic spectral power. is required. Next, regarding the logarithmic spectral power for each band, in order to make it appear as the same feature parameter regardless of whether the input voice is a loud voice or a soft voice, the following rough (transformation is performed) is performed. .

First, the output value of the logarithmic conversion unit 8-21, that is, the logarithmic spectral power of the highest frequency band, is multiplied by a weighting constant, which has been stored in the constant storage unit 9 in advance, by a multiplier 10. ”1. As mentioned above.

Since the bandpass filter 4-11 has a wider bandwidth than the other bandpass filters 4-1 to 4-(?!-1'), the weight for multiple channels can be calculated using ]channel. 8.If one channel of the highest frequency band corresponds to the bandwidth of three channels of low frequencies,

For weighting, "3" is stored as a constant in the constant storage section 9, and 2
The multiplier 10 converts the output value of the logarithmic conversion unit 8-22 into 3
It will be doubled.

The average value calculation unit 11 calculates an average value of the logarithmic spectral power by band in which the weighting described above is taken into consideration. For example, the output values of the logarithmic conversion units 8-1 to 8-71 are P;, P2', aT, pn'-1, p;, respectively.

Assuming that the weighting constant is W, the average value F is as follows.

n-1-4-W subtracters 12-1 to 1'2-(72-1) are provided corresponding to logarithmic conversion units 8-1 to 8-(n-1). That is, there is no need to provide a subtracter corresponding to the logarithmic conversion section 8-H.
The logarithmic spectral power P2'Z for each band is.

Used only for calculating the average value] 1. After calculating the average value, it is removed. Subtractor 12-1-12-(tl-1)
is the logarithmic spectral power P,′,P; for each band.

. . . The output P of the average value calculation unit 11 is subtracted from Pn. That is, subtracters 12-1 to 12-C11-
The outputs ξ of 1) are as follows.

This subtracter 12-1 to 12-(n-1)(7) output P.

is output to the speech recognition unit 13 as a feature parameter for verification.

The speech recognition unit 13 uses a feature parameter time series consisting of a set of Cn-1) feature parameters.

The input speech is recognized by comparing it with a standard feature parameter time series registered in advance in the dictionary 10 using, for example, the DP matching method. That is, to put it simply, the time axis is normalized, and the distance (
Standard 4 where the result of adding Pi-Pi) from 1-1 to i-m and adding this for a series of time series is the minimum
A recognition result is a monosyllable or a word corresponding to the characteristic parameter.

In order to test the effect of the present invention, the present inventors 1.

The following experiment was conducted. First, from the first channel
Up to 9 channels, 180KHz to 7.8 IUl
A 19-band filter that covers the band up to Z was prepared. Especially the 17th channel and the 18th channel.

Describe the characteristics of the 19th channel bandpass filter.

The center frequencies are 5145Hz and 5910Hz, respectively.
, 7020Hz.

The lower limit frequency is 4800Hz, 551411z,
6334 Hz, upper limit frequency is 5514 Hz, 63
There are 34h and 7800H7te, each with a bandwidth of 71
4 Hz, 820 Hz, and 1466 Hz. Then, according to the conventional method, speech recognition was performed using 19 normalized spectral powers obtained by normalizing the band-specific logarithmic spectral powers of all channels as feature parameters.

Next, from the 1st channel to the 16th channel.

Prepare the same bandpass filter as above, and replace the bandpass filter from the 17th channel to the 19th channel, and set the lower limit frequency to 4800FI7. , upper limit frequency is 7, 8
Using a KH7 bandpass filter, spectrum analysis is performed using 17 channels as explained in the above example, and the 2-weighting constant is set to 3, and the output value of the 17th channel is weighted to calculate the average value. Then, the average value was subtracted from the 16 band-specific logarithmic spectra excluding the output value of the 17th channel, and the result was used as the characteristic parameter. Based on these 16 feature parameters, speech recognition was performed by comparing them with standard feature parameters consisting of a newly created set of 16 feature parameters.

Regarding the speech recognition rate, results similar to those obtained using the above 19 feature parameters were obtained.

It should be noted that previous experiments have shown that if only spectrum analysis from the 1st channel to the 16th channel is performed from the beginning, the speech recognition rate will decrease because high-frequency information is not taken into account at all. There is.

We further narrowed down the experiment by changing one frequency band, but were able to obtain similar effects.

(D As described in detail, according to the present invention, by simple means,
(reducing the speech recognition rate), the amount of matching/storing feature parameters can be reduced.

It is possible to reduce the amount of memory, calculation mechanism, etc., and by consolidating the high frequency band parts, it is possible to reduce the amount of hardware required for spectrum analysis. By further widening the nine frequency bands, it becomes possible to improve the speech recognition rate.

[Brief explanation of drawings]

The figure shows the configuration of an embodiment of the present invention. In the figure, 1 is a voice input section, and 2 is a parameter extraction section. 3 is a spectrum analysis section, 4-1 to 4-F1 are band pass filters, and 5-1 to 5-21 are DH, devices. 6-1 to 6-n are analog-to-digital conversion circuits;
7 is a spectrum normalization unit, 8-1 to 8-n are logarithmic conversion units, 9 is a constant storage unit, 10 is a multiplier, 11 is an average value calculation unit, and 12-1 to 12-(n −1) are subtraction units. vessel, 1
3 represents a speech recognition unit, and 14 represents a resignation letter. Patent applicant Fujitsu Limited

Claims (3)

[Claims] (1) In a speech recognition method that recognizes speech by collating feature parameter time series obtained by frequency analysis of speech, - spectrum analysis is performed over a wide speech frequency band, and high frequency band parts of the analysis results are separated. A speech recognition method characterized in that after obtaining a normalized spectrum by weighting the above, the spectrum from which the high frequency band portion has been removed is used as a feature parameter time series for comparison. (2) The above weighting is normalized by 9 After logarithmically transforming the analyzed spectrum, weighting the high frequency band part and finding the average value, and correcting the sum of the spectra based on the average value is a predetermined value such as zero. A speech recognition system according to claim (1), characterized in that: (3) The above spectral analysis uses multi-channel bandpass filters and converts the output of each filter into band-specific spectral power. After converting to obtain the logarithmic spectral power for each band, weight one channel of the highest frequency band, and obtain the average value for weighting of all channels. The voice recognition method according to claim 1, wherein the weighting is performed by subtracting the average value from one channel of the highest frequency band (using spectral power).