JPH0146079B2 - - Google Patents

Description

[Detailed Description of the Invention] (A) Technical Field of the Invention The present invention relates to a speech recognition device, particularly a speech recognition device 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 speech recognition device that can reduce the amount of parameters in a time series of feature parameters to be compared and reduce the amount of analysis hardware without reducing the speech recognition rate.

(B) Technical background and issues As a speech recognition method, in order to perform wideband speech frequency analysis, bandpass filters with multiple channels are used, and the output of each filter is converted to band-specific spectral power by rectification and integration. Then, logarithmically transform them to obtain the logarithmic spectral power for each band. To normalize the spectrum, transform the logarithmic spectral power for each band so that the average value of all channels is zero, and then calculate all normalized log spectral powers. The logarithmic spectral power for each band is used as a feature parameter time series for matching, and it is matched with a standard feature parameter time series registered in a dictionary in advance using, for example, a dynamic programming (DP) matching method, to identify monosyllables, words, etc. A method for performing voice recognition is known.

In the above speech recognition method, in order to increase the speech recognition rate, it is necessary to increase the number of band 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 memory required can be reduced, but the speech recognition rate will deteriorate.

By the way, the inventors of the present invention have accumulated many experiments and researches prior to completing the present invention, and have discovered the following characteristics in speech recognition. Better results can be obtained when audio frequency analysis is performed over a wide band, including the high frequency band, but especially for the high frequency band, the relative amount of audio energy in the power spectrum for each sampling is For example, whether the peak of the power spectrum is in the 5KHz frequency region or the 7KHz frequency region is not that important for speech recognition. this is,
This is probably because the human ear has difficulty perceiving slight differences in frequency 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. When we conducted an experiment, we found that the recognition rate did not decrease compared to when parameters for all channels, including the high frequency part, were used as the feature parameter time series. On the other hand, it has been found that if the high frequency band portion is not added to the normalization conditions from the beginning, the speech recognition rate decreases.

(C) Object 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 matched without reducing the speech recognition rate, thereby reducing the amount of memory, etc. The aim is to reduce the amount of hardware required for spectrum analysis. In other words, the objective is to further improve the speech recognition rate with the same amount of feature parameters as before. Therefore, the speech recognition device of the present invention is a speech recognition device that recognizes speech by collating feature parameter time series obtained by frequency analysis of speech. and weighting means for weighting the high frequency band portion of the analysis result by the spectrum analysis section according to its bandwidth, and average value calculation means for calculating the average value of the analysis results including the weighted results, and and a spectrum normalization unit that normalizes the spectrum excluding the high frequency band portion based on the value, and uses the parameter time series of the normalized spectrum from which the high frequency band portion is removed as the feature parameter time series for matching. It is characterized by the fact that Hereinafter, embodiments will be described with reference to the drawings.

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

In the figure, 1 is an audio input section, 2 is a parameter extraction section, 3 is a spectrum analysis section, 4-1 to 4-n
is a band pass filter, 5-1 to 5-n are rectifiers, 6-1 to 6-n are analog-to-digital conversion circuits, 7 is a spectrum normalization section, 8-1 to 8-n are logarithmic conversion sections, and 9 is a Constant storage unit, 10 is a multiplier, 11 is an average value calculation unit, 12-1 to 1
2-(n-1) is a subtracter, 13 is a speech recognition unit, 1
4 represents a dictionary.

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 weights and normalizes the outputs of the spectrum analysis unit 3 that performs spectrum analysis over a wide audio frequency band and the spectrum analysis unit 3 with respect to the high frequency band portion. Feature parameters for matching the normalized spectra excluding P ₁ , P ₂ ,...
It has a spectrum normalization unit 7 that outputs as P _o-1 .

The spectrum analysis section 3 includes a plurality of (n pieces) for each band.
It has band pass filters 4-1 to 4-n. In the figure, the upper bandpass filter 4-1
The passing frequency increases as one goes to the lowest level. Bandpass filters 4-1 to 4-n
are arranged so that, for example, the 3 dB attenuation points of adjacent band pass filters coincide, 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) are set to have a bandwidth of, for example, 170Hz to 620Hz, but the bandpass filter 4-n, which has the highest frequency band, has a wide band width of, for example, 3KHz. It is designed to have band characteristics.

The audio signal from the audio input section 1 is filtered by band by bandpass filters 4-1 to 4-n,
The signals are respectively input to rectifiers 5-1 to 5-n. Each of the rectifiers 5-1 to 5-n performs rectification and smoothing of the input signal with a rectification and integration time constant of, for example, 10 ms. The outputs of the rectifiers 5-1 to 5-n are analog
The signal is input to digital converters 6-1 to 6-n, and the spectrum power for each band is expressed as a digital quantity. The conversion result is output to the spectrum normalization section 7.

The spectrum power for each band inputted to the spectrum normalization unit 7 is logarithmically transformed by the logarithmic conversion units 8-1 to 8-n so that the output value is proportional to the strength of the sound perceived by humans, and is converted into a logarithm for each band. The spectral power is determined. Next, in order to make the input voice appear as the same feature parameter regardless of whether the input voice is a loud voice or a soft voice, the following conversion is performed on the logarithmic spectral power for each band.

First, a weighting constant stored in the constant storage section 9 in advance is applied to the output value of the logarithmic conversion section 8-n, that is, the logarithmic spectral power of the highest frequency band.
Multiply by 0. This is because, as mentioned above, the bandpass filter 4-n has a wider bandwidth than the other bandpass filters 4-1 to 4-(n-1). This is because it has the weight of the channel. If one channel of the highest frequency band is
When the bandwidth corresponds to three channels in the low frequency band, "3" is stored as a weighting constant in the constant storage section 9, and the output value of the logarithmic conversion section 8-n is tripled by the multiplier 10. will be done.

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, if the output values of each logarithmic conversion unit 8-1 to 8-n are P' ₁ , P' ₂ , ..., P' _o-1 , P' _o , and the weighting constant is ω, then the average The value P is as follows.

The subtracters 12-1 to 12-(n-1) are provided corresponding to the logarithmic conversion units 8-1 to 8-(n-1). That is, a subtracter corresponding to the logarithmic conversion section 8-n is not provided, and the logarithmic spectral power P′ _o for each band is
is used only for calculating the average value, and is removed after calculating the average value. Subtractors 12-1 to 12-
(n-1) is the logarithmic spectral power P′ ₁ for each band,
The output of the average value calculation unit 11 is subtracted from P′ ₂ , . . . , P′ _o-1 . That is, subtracters 12-1 to 12
The outputs P _i of -(n-1) are each as follows.

P _i = P' _i - (i=1, 2,..., n-1) Output of this subtractor 12-1 to 12-(n-1)
P _i is output to the speech recognition unit 13 as a matching feature parameter.

The speech recognition unit 13 collates the feature parameter time series consisting of a set of (n-1) feature parameters with the standard feature parameter time series registered in advance in the dictionary 10 by, for example, the DP matching method. Performs input voice recognition. That is, to put it simply, the time axis is normalized, and the distance (P _i - P' _i ) ² between the m input feature parameters P _i and the standard feature parameter P' _i at the corresponding time point is defined as i=
1 to i=m, and the monosyllable or word corresponding to the standard feature parameter for which the result of adding this for a series of time series is the minimum is set as the recognition result.

The present inventors conducted the following experiment in order to test the effects of the present invention. First, we prepared 19 band filters covering the band from 180KHz to 7.8KHz from the 1st channel to the 19th channel. In particular, the characteristics of the band filters for the 17th, 18th, and 19th channels are as follows: The center frequencies are 5145Hz, 5910Hz, and 7020Hz, the lower limit frequencies are 4800Hz, 5514Hz, and 6334Hz, and the upper limit frequency is 5514Hz.
Hz, 6334Hz, 7800Hz, and the bandwidth is respectively
They are 714Hz, 820Hz, and 1466Hz. Then, using 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 band filter as above, replace the band filter from channel 17 to channel 19, and set the lower limit frequency to 4800Hz and the upper limit frequency to 7.8K.
Using a Hz band filter, perform spectrum analysis on 17 channels as explained in the above example, set the weighting constant to "3", weight the output value of the 17th channel, and calculate the average value. death,
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, but the speech recognition rate was
Similar results were obtained when using 19 feature parameters.

In addition, 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 information in the high frequency region is not taken into account at all. There is.

Furthermore, we repeated the experiment by changing the frequency band and were able to obtain similar effects.

(E) Effects of the Invention As explained above, according to the present invention, the amount of matching/storing feature parameters can be reduced by simple means without reducing the speech recognition rate, and the amount of memory and calculation mechanism can be reduced. By consolidating the high frequency band parts, the amount of hardware for spectrum analysis can be reduced. moreover,
By widening the frequency band, 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 an audio input section, 2 is a parameter extraction section, 3 is a spectrum analysis section, 4-1 to 4-n
is a band pass filter, 5-1 to 5-n are rectifiers, 6-1 to 6-n are analog-to-digital conversion circuits, 7 is a spectrum normalization section, 8-1 to 8-n are logarithmic conversion sections, and 9 is a Constant storage unit, 10 is a multiplier, 11 is an average value calculation unit, 12-1 to 1
2-(n-1) is a subtracter, 13 is a speech recognition unit, 1
4 represents a dictionary.

Claims (1)

[Scope of Claims] 1. A speech recognition device that recognizes speech by collating feature parameter time series obtained by frequency analysis of speech, comprising: a spectrum analysis unit that performs spectrum analysis over a wide speech frequency band; Weighting means for weighting the high frequency band portion of the analysis result by the spectrum analysis section according to its bandwidth; and average value calculation means for calculating the average value of the analysis results including the weighted results, and based on the average value. and a spectrum normalization unit that normalizes the spectrum excluding the high frequency band portion, and uses the parameter time series of the normalized spectrum from which the high frequency band portion is removed as the feature parameter time series for matching. A voice recognition device featuring: 2 The spectrum analysis section is configured to use multi-channel bandpass filters and convert the output of each filter into spectrum power for each band, and the spectrum normalization section logarithmically transforms the spectrum power for each band. After determining the logarithmic spectral power for each band, weight one channel of the highest frequency band to determine the weighted average value of all channels,
The speech recognition device according to claim 1, wherein the speech recognition device is configured to output spectral power excluding one channel in the highest frequency band, which is obtained by subtracting the weighted average value from the logarithmic spectral power for each band. .