JPS61281300A - Voice recognition equipment - 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 [Field of Industrial Application] The present invention relates to a speech recognition device, and particularly to a technique for normalizing sound source characteristics to remove fluctuation factors due to differences in power level and speakers.

[Summary of the invention]

In this invention, a sound spectrum is obtained from input speech by an acoustic analysis section, a sound source spectrum characteristic is obtained from this sound spectrum using the least squares method, and the obtained sound source spectrum is subtracted from the original sound spectrum to normalize the sound source characteristic. In a device that obtains a converted speech spectrum and performs speech recognition using it, even if the frequency axis is not divided at equal logarithmic intervals in the acoustic analysis section, it is equivalent to dividing it at equal logarithmic intervals. This system is equipped with a correction means to obtain a speech spectrum of
, the fluctuation components of the sound source characteristics are further reduced, and the performance of the speaker-independent speech recognition device can be improved.

[Conventional technology]

Speech is a phenomenon that changes along the time axis, and has a spectrum.
Unique words and words are created by uttering sounds with ever-changing patterns. Speech recognition is a technology that automatically recognizes words and phrases spoken by humans.
At present, it is extremely difficult to achieve speech recognition that is comparable to the human auditory function. For this reason, most speech recognition methods currently in practical use are performed by pattern matching a standard pattern of a word to be recognized and an input cover pattern under certain conditions of use.

By the way, in the case of unspecified speaker input in which the speaking speaker input to the speech recognition device is not limited to registered speakers, the method of speaker normalization is an important issue.

FIG. 2 is a block diagram of an example of a speech recognition device that also takes this into account, in which speech input from a microphone (1) is supplied to an acoustic analysis circuit (2). This acoustic analysis circuit (2) extracts acoustic parameters representing the characteristics of the input speech pattern. Various acoustic analysis methods can be considered to extract these acoustic parameters, but one example is to use a bandpass filter and a rectifier circuit as one channel, and to divide such a channel into multiple channels each having a passband that is obtained by dividing the audio band. A method is used to extract temporal changes in the spectral pattern as a way of arranging the bandpass filters.

That is, in the acoustic analysis circuit (2), the audio signal from the microphone (1) is transmitted to the amplifier (21) and the band -
The signal is supplied to an A/D converter (23) via a limiting low-pass filter (22) and converted into a 12-bit digital audio signal at a sampling frequency of, for example, 12.5 kHz. This digital audio signal is processed by the digital bandpass filters (24r) (242) of each channel of a 16-channel bandpass filter bank, for example.
..., (24te) is supplied. These digital band pass filters (241) and (242).

..., (24xs) is, for example, Butterworth 4
Consists of the following digital filters, from 250Hz to 5
．． Each band obtained by dividing the band up to 5 kHz at equal intervals on the logarithmic axis becomes the pass band of each filter.

In other words, the frequency is divided into 1 at equal intervals on a log scale.
A bandpass filter bank for six channels is configured.

Each digital band pass filter (241), (2
42).

The output signals of ..., (241s) are respectively connected to rectifier circuits (25z), (252), ...,
(25ts) and these rectifier circuits (25ts).
1), (252), ... (251g) outputs are each digital low-pass filter (261)
.

(262), ..., (261J). These digital low-pass filters (261),
(262).

..., (261g) is composed of, for example, an FIR low-pass filter with a cutoff frequency of 52.8 Hz.

Each digital low-pass filter C261), (2B2), ... which is the output of the acoustic analysis circuit (2)
The output signal of (261g) is supplied to the sampler (3) that constitutes the feature extraction circuit.
), the digital low-pass filter (261), (
262) ,...

(261s) output signal with a frame period of 5.12m5e
Sample every c. Therefore, from this, the sample time series At(nl (i = 1. 2. ”I6
;n is the frame number and n=1.2.・・・
..., N) is obtained.

The output from this sampler (3), that is, the sample time series At of the audio spectrum (nl is the sound source information normalization circuit (
4), thereby eliminating differences in vocal cord sound source characteristics and differences in utterance loudness between speakers of the speech to be recognized.

Conventionally, a method has been proposed that normalizes the loudness of vocalizations and normalizes individual differences in vocal cord sound source characteristics by using a least squares approximation straight line of the voice spectrum expressed logarithmically on both the amplitude and frequency axes. (Reference: "A method of word speech recognition using nonlinear spectral matching" Miwa, Kokan, Makino, Kido; Transactions of the Institute of Electronics and Communication Engineers '81/I Vol.
J64-D No, I P46-P53).

That is, the sound source information normalization circuit (4) uses the sample time series A supplied from the sampler (3) for each frame period.
At(nl=log(Ai(n)+B) for t(n)
...(1) A logarithmic transformation is performed. This (1
), B is set to a value such that the noise level is hidden by the bias.

Here, the vocal cord sound source characteristics are modeled using the equation yi =at+b.

At this time, yt representing vocal cord sound source characteristics is expressed as voice spectrum x
Find coefficients a and b as least squares approximation straight lines of 1.

That is, coefficients a and b are determined by the following equations.

(T=16)...(2) (1=16)...(3) Then, the normalized audio spectrum 2L is given by the following equation.

zI=xI-yL...(4) However,
Here, considering the difference between unvoiced sounds and voiced sounds, when the slope of the least squares approximation straight line is a≧0, only the voice level is normalized, and the overall slope of the spectrum is not changed.

That is, in the case of voiced sounds, a<0 as shown in Figure 3.
On the other hand, in the case of unvoiced sounds, a≧0 as shown in FIG. Therefore, in the case of unvoiced sounds, only the voice level is normalized without removing the slope of the vocal cord sound source spectrum characteristics.

Putting the above in order, we get the following.

Letting the normalized parameters of the sound source be Pi(nl),
When a(n)<Q, the parameter Pi(n) is Pi(n)
=Ai(n) (a(n) ・i + b(n))
...It is expressed as (5).

Also, when a (fil≧0, only the level is normalized, and the parameter Pi(n) is the acoustic parameter time series P i (n) in which the differences in the vocal cord sound source characteristics are normalized and removed, and the sound source information is normalized. obtained from the conversion circuit (4).

Acoustic parameter P from this sound source information normalization circuit (4)
i(n) is supplied to the intra-speech interval parameter memory (5). In this voice section parameter memo (January 5), upon receiving the voice section judgment signal from the voice section judgment circuit (6),
A parameter pt(nl is stored for each determined voice section.

The voice section determination circuit (6) is a zero cross counter (61)
, power calculation circuit (62), and voice section determination circuit (63)
A digital audio signal from the A/D converter (21) is supplied to a zero cross counter (61) and a power calculation circuit (62). Zero cross counter (61
) counts the number of zero crosses of the 64 samples of digital audio signal within this one frame period every 5.12 m5ec, and the count value is the first input of the audio section determination circuit (63). Supplied at the end. The power calculating circuit (62) calculates the power of the digital audio signal within one frame period, that is, the sum of squares, for each frame period, and the output power signal is sent to the audio section determining circuit (63).
) is supplied to the second input of the circuit. Voice section determination circuit (6
3) is further supplied with the values of sound source normalization information a and b from the sound source information normalization circuit (4) to its third input terminal. Then, in this speech section determining circuit (63), the number of zero crossings, the power within the section, and the sound source normalization information are processed in a composite manner, and a process for determining silence, unvoiced sound, and voiced sound is performed, and a speech section is determined.

The voice interval determination signal indicating the determined voice interval from the voice interval determination circuit (63) is supplied to the voice interval parameter memo (January 5) as an output of the voice interval determination circuit (6), and the voice interval determination signal indicating the determined voice interval is This memory (5
) is the normalized acoustic parameter time series Pi of the sound source information
(n) is stored.

The acoustic parameter time series Pi(n) thus stored in the memo (January 5) is supplied to the speech recognition section (7) and subjected to speech recognition.

That is, for example, it is registered in advance! ! ! The distance between the target word and the acoustic parameter time series is calculated,
The word with the smallest calculated distance will be recognized as the speech recognition output.

In this case, for time normalization processing, so-called DP matching using dynamic programming may be performed in distance calculation.

In addition, as this time normalization process, by estimating the trajectory drawn by the acoustic parameter time series P 1 (n) in its parameter space and resampling along this trajectory,
A new parameter time series Qi that is time-normalized from the acoustic parameter time series Pi(n) is formed, and a standard pattern of the recognition target word is registered in the form of this new parameter time series Qi(e). In addition, this parameter Qi (
E) may be used to calculate the distance during recognition (for example, see Japanese Patent Application No. 59-106177) [Problems to be Solved by the Invention] In the case of the conventional speaker normalization method as described above, It uses the fact that the sound source spectrum characteristic yt can be obtained by linear approximation when the frequency axis is divided at equal logarithmic intervals, so when using a bandpass filter bank for acoustic analysis, it is possible to divide the frequency axis at equal logarithmic intervals. There was a constraint that each channel had to be defined.

On the other hand, as a band-pass filter bank, the MET filter has a frequency axis that corresponds to the human auditory characteristics.
There is also a request to use a scale.

In addition, when a multi-channel bandpass filter bank is constructed by dividing the audio frequency band at equal logarithmic intervals, the disadvantages that the low frequency side becomes finer, making filter design difficult and increasing the number of channels can be eliminated. For,
The frequency axis takes advantage of the advantages of the mel scale and the logarithmic axis, such as dividing the frequency bands at equal intervals on the mel scale and taking each channel on the low frequency side, and leaving them at equal logarithmic intervals on the high frequency side. There is also a demand for configuring a bandpass filter bank.

However, as mentioned above, the conventional method assumes a logarithmically evenly spaced band-pass filter bank, so it is not possible to use the sound source information normalization method using the least squares method, which is easy to calculate, as described above. To none.

Furthermore, in the case of the conventional normalization method described above, voiced sounds and unvoiced sounds are separated, and in the case of unvoiced sounds, the slope of the sound source spectrum is not removed.

This is because the phonological characteristics of voiced sounds are determined by the position of the formants and are not related to the spectral slope, whereas the characteristics of unvoiced sounds are said to be determined not only by formants but also by the spectral slope. to cause.

However, the slope of the spectrum is expected to vary considerably depending on the speaker and the way they speak. Therefore, in the conventional case, such large fluctuation factors are still included in the sound source information.

[Means for solving problems]

The present invention has been made to solve the above-mentioned problems, and includes an acoustic analysis means for determining the audio spectrum of an input audio signal, and an acoustic analysis means that calculates the audio spectrum of an input audio signal so that the frequency axis of the audio spectrum is equal to that when divided at equal logarithmic intervals. a correction means for performing correction, a means for obtaining a sound source spectrum characteristic from the voice spectrum as a least square approximation straight line, a subtraction means for subtracting the sound source spectrum from the corrected voice spectrum, and a normalization of the sound source information from the subtraction means. and means for performing speech recognition using the converted speech spectrum.

[Effect]

Even if the bandpass filter bank is not divided into frequency bands at equal logarithmic intervals and the center frequency of each channel is determined, it is corrected by the jli positive means so that it is equal to the frequency band divided at equal logarithmic intervals.

Therefore, the sound source information can be normalized using the sound source spectrum obtained by the least squares method on the sound spectrum.

In addition, in the subtraction means, the sound source spectrum is subtracted from the voice spectrum without distinguishing between voiced and unvoiced sounds, and the slope of the sound source spectrum is removed, making it possible to further remove variations in sound source information including differences in speakers. .

〔Example〕

FIG. 1 is a block diagram of an embodiment of the present invention, in which audio input from a microphone (1) is supplied to an acoustic analysis section (2) using a band-pass filter bank, and the analysis output is sent to a sampler (3). from which the acoustic parameters are obtained.

This acoustic parameter is supplied to the correction circuit (41) of the sound source information normalization circuit (4).

In this case, the sound source power spectrum at frequency f is set to yt, and the sound source characteristics are modeled using the following function.

log yt =C1og rt +a 1 ・
-(7) This model means that when the center frequency and bandwidth of the bandpass filter are set at equal intervals on the log scale, the source logarithmic power spectrum is expressed as a straight line. At this time, if the least squares approximation straight line of the audio logarithmic power spectrum is taken as the sound source power spectrum, then
The sound source characteristics are determined by the least squares method.

In this way, the frequency axis of the logarithmic power spectrum is
When the scale is used, the sound source characteristics can be found by simply applying the method of least squares, but when the scale is not log scale, correction is required.

This correction is performed in the correction circuit (41). The following three types of correction can be considered as the correction performed by this correction circuit (41).

■ In the case of a bandpass filter bank, bandwidth correction ■ Conversion of the frequency axis to log scale ■ Weight correction for errors when evaluating the sum of squared errors The first bandwidth correction is as follows. It will be done.

The center frequency of the i-channel bandpass filter is fI
, the lower limit frequency of the band is F(-1, the upper frequency limit is Ft, the bandwidth is ΔF", then ΔF[-FL F[-1...(8) The average output power of the i channel is Xi, and xl is When a signal with a flat spectrum is input as a standard and the output of each channel is corrected to be constant, the corrected output x
I' is log x+'=1og x(-1og (
ΔF+/ΔFt)・・・aω Here, log xl'=Xt', log xt=X
If t +log (ΔFL/ΔFt)=DF+, then XI'=xL -DFI ・ ・
- (11), and the output level of the bandpass filter is corrected.

The conversion of the second frequency axis to the log scale is done by the following equation.

hI=logft (12) Next, weight correction for the third error is performed as follows.

When the frequency axis of the audio power spectrum is not plotted on a log scale, the sample points of the audio spectrum are arranged at uneven intervals on the log scale, creating the effect that they are arranged at equal intervals on the log scale. Therefore, a weight ω' is applied to the squared error of each spectrum value.

Here, if Yt - 1og yt, the sum of squared errors e is The weight ω1 is determined as follows.

Δht=1og F(-1og Ft-t...
(14) If the bandwidth is Δg at equal intervals on the log scale, then (Δg - ΔhN) each part of the audio spectrum corrected as above is supplied to the sound source spectrum characteristic extraction circuit (42), minimum 2
An approximate straight line can be found by multiplication.

That is, by substituting equations (7) and (12) into equation (13), from a e / a c = 0... (17) from a e / a d = Q... (1g) Therefore , ... (19) ... (20) Slope C2 intercept d of equation (19), equation (20) to equation (7)
Find the value of

The sound source spectrum characteristics are expressed by c and d.

Each part of the sound source spectrum obtained in this way is subtracted by a subtraction circuit (4
3) and is subtracted from each part of the audio spectrum from the correction circuit (41). In other words, if the logarithmic power spectrum normalized to the sound source information is Z[, then the subtraction circuit (
43), the following calculation is performed.

Z+ =Xt'-Yt =Xt' (c, hl +d)...(
21) In this way, the subtraction circuit (43) obtains the sound power spectrum Zi whose sound source information has been normalized.

This is then supplied to the speech recognition unit (7), where distance calculation is performed on the standard pattern of the recognition target word registered as described above, and a recognition output is obtained by determining the minimum value of the calculated distance.

In this case, as mentioned above, so-called DP matching may be performed, or the trajectory drawn by the acoustic parameter in the parameter space may be estimated, the trajectory may be resampled to obtain new parameters, and time normalization may be performed. may be used to calculate the distance to the standard pattern.

〔Effect of the invention〕

According to this invention, even when the frequency axis is not arranged at equal logarithmic intervals during acoustic analysis, the sound source spectrum is determined by the simple least squares method, and the sound source spectrum is subtracted from the original sound spectrum to be removed. Information normalization methods can be used.

Therefore, when using a bandpass filter bank in the acoustic analysis section, the passing frequency of each channel can be designed on various frequency axes suitable for speech recognition, which increases the degree of freedom in design. Become.

In addition, in this invention, the slope of the sound source spectrum is removed not only for voiced sounds but also for unvoiced sounds, so considering that the slope of the sound source spectrum for unvoiced sounds varies depending on the speaker, the sound source characteristics The fluctuation component of is further reduced, and the performance of the speech recognition device compatible with unspecified speakers can be improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of the main part of this invention, and FIG.
The figure is a block diagram showing the configuration of an example of a speech recognition device.
4 and 4 are characteristic diagrams for explaining sound source spectral characteristics. (2) is the acoustic analysis section, (4) is the sound source information normalization circuit,
(4) is a correction circuit, (42) is a sound source spectrum characteristic extraction circuit, and (43) is a subtraction circuit.

Claims (1)

[Scope of Claims] (a) Acoustic analysis means for determining the audio spectrum of an input audio signal; (b) Correction means for making corrections so that the frequency axis of the audio spectrum is equal to that when divided at equal logarithmic intervals. (c) means for determining a sound source spectrum characteristic as a least squares approximation straight line from the sound spectrum; (d) a subtraction means for subtracting the sound source spectrum from the corrected sound spectrum; and (e) a sound source from the subtraction means. A speech recognition device comprising means for performing speech recognition using a normalized speech spectrum of information.