JP2954591B2 - Speech feature extraction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention enables stable speech recognition with a high recognition rate irrespective of a speaker by extracting a feature amount related to human articulation from input speech. The present invention relates to an audio feature extraction method.

<Conventional technology> Conventionally, input speech is recognized by a speech recognition device as follows. That is, a speech feature extraction device that passes an input speech through several BPFs (bandpass filters) and extracts an output value as a feature of the input speech extracts a feature of a word to be recognized in advance. Some quantities are stored in the form of time series as standard patterns. Then, when an unknown voice is input, the feature amount of the unknown input voice is extracted as described above, and the similarity between the feature amount of the unknown word and the feature amount of the standard pattern is checked.
The words are identified and determined based on the similarity.

At this time, the similarity is represented by the distance between the time-series pattern of the feature amount of the unknown word and the standard pattern, and the distance measure is the difference in the logarithm of the output value of each BPF or its squared value. In many cases, a value added (averaged) over a band is used. In addition, as the above-mentioned feature quantity, the cepstrum coefficient, the ZCC (number of zero crossings) of the output of each BPF, or the ZCI
The value of (zero-crossing interval) may be used.

<Problems to be Solved by the Invention> However, the above-described conventional speech recognition apparatus uses the feature amount directly related to the spectrum of the input speech to determine the similarity (distance scale) between the standard pattern and the pattern of the feature amount of the input speech. ) Is calculated based on the difference between the output values of each BPF. Therefore, depending on the physiological difference of speakers (difference in vocal tract length, etc.), the spectrum of the generated speech may be different even if the same word is generated. In many cases, there is a problem that erroneous recognition or rejection may occur, and high recognition performance cannot be obtained.

This is because the output value of each BPF is used as it is as the feature value, and it is difficult to extract essential information necessary to identify each phoneme, affected by the fluctuation of the spectrum by the speaker. That's why. That is, in order to effectively identify phonemes, it is important to convert the output of each BPF into a feature quantity related to human articulation and use it.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech feature extraction method capable of stably and correctly recognizing speech without being affected by spectrum fluctuations caused by a speaker.

<Means for Solving the Problems> In order to achieve the above object, an invention according to claim 1 is a voice feature extraction method for performing frequency analysis of an input voice and extracting a voice feature from an obtained spectral component. A first step of obtaining a discriminative feature quantity representing a degree of a discriminative feature representing frictional property based on the spectrum component obtained by the frequency analysis; and A second step of obtaining a discriminative feature representing the degree of the discriminating feature representing voiced, and a discriminating feature representing the flatness of the spectrum by nasalization based on the spectral component obtained by the frequency analysis. A third step of obtaining a discriminative feature amount representing the degree, and a degree of the discriminative feature representing the buzziness based on the step component obtained by the frequency analysis. A fourth step of obtaining a discriminative feature amount representing the matching, and a discriminative feature amount representing the degree of the discriminative feature representing the flatness of the spectrum of the middle frequency band based on the spectrum component obtained by the frequency analysis. Based on the fifth step to be determined and the spectral components obtained by the frequency analysis,
The method includes a sixth step of obtaining a discriminating feature amount indicating a degree of the discriminating feature indicating the dispersion of the spectrum, and extracting a feature amount of a sound related to human articulation.

<Operation> A voice is input, and the input voice is subjected to frequency analysis. Then, based on the spectral components obtained by the frequency analysis, in the first step, a discriminative feature amount indicating the degree of the discriminative feature indicating the frictional property is obtained by a procedure corresponding to the discriminative feature. Hereinafter, similarly, in the second step, a discriminative feature amount indicating the degree of the discriminative feature indicating voiced is obtained, and in the third step, a discriminative feature amount indicating the flatness of the spectrum due to nasalization is obtained. , In the fourth step, a discriminative feature representing the buzziness is obtained, in the fifth step, a discriminative feature representing the flatness of the spectrum of the middle frequency band is obtained, and in the sixth step, the dispersion of the spectrum is obtained. Is obtained. In this way, the above-mentioned six types of feature amounts related to human articulation, which are discriminative feature amounts, are extracted.

<Example> Hereinafter, the present invention will be described in detail with reference to an illustrated example.

FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention. The audio signal input from the microphone 1 is amplified by the amplifier 2 and input to the feature extraction unit 3. The feature extraction unit 3 includes a plurality of channels (hereinafter referred to as ch.)
(16 channels in this embodiment). The center frequency of each of the BPFs is 300 Hz for the first channel and 3400 Hz for the 16th channel, and is set so that the interval between them is equal. FIG. 2 shows an outline of the frequency characteristics of each BPF.

For the audio signal input to the feature extraction unit 3 configured as described above, the power as a feature amount is calculated for each unit time (frame) by each BPF. At that time, ZCC or ZCI may be used as the feature amount instead of the power. The feature amount obtained for each of the above BPFs corresponds to an approximate spectrum of an input voice. Hereinafter, this feature amount is defined as hi (i = 1 to 1).
6: ch. Number). In this way, the feature amount obtained for each BPF is input to the discriminative feature extraction unit 4. The discriminative feature extraction unit 4 is a main part of the present invention, and extracts a discriminative feature representing the degree of a discriminative feature related to human articulation from the feature hi as described later in detail. This is the part to do.

The time series of the discriminative feature amounts of the input speech obtained for each frame by the discriminative feature amount extraction unit 4 is input to the word recognition unit 5 and is used to discriminate the standard patterns stored in the standard pattern storage unit 6. The degree of similarity between the feature amount and the time series is calculated. At this time, the similarity calculation is performed by a DP matching method or the like. Then, the speech input word recognized based on the result of the similarity calculation is displayed on the result display unit 7.

Next, a method of calculating a discriminative feature amount representing the meaning and the degree of the discriminative feature will be described in detail.

The above distinctive features are related to human articulation,
There are six types as follows.

“Strident” is a distinctive feature that indicates strong friction,
It can be determined based on the degree of concentration of energy in the high frequency band (2500 Hz).

“Voiced” is a distinctive feature that indicates the presence of vocal cord vibration, and can be almost determined by the presence of energy in the 120 Hz to 750 Hz band.

“Nasal” is a distinctive feature that indicates the degree of nasalization,
As the expiration also flows into the nasal cavity, the determination can be made based on the degree to which the spectrum of the first formant region becomes flat by the pole-zero pair.

“Murmur / buzz” is a distinctive feature that indicates the degree of buzziness, and can be determined by the degree of concentration of energy in the low frequency band.

“Diffuse” is a discriminating feature indicating that the spectrum is relatively flat, and can be determined by the degree of flatness of the spectrum in the middle frequency band to the high frequency band (800 Hz or more).

“Compact” is a distinctive feature that indicates whether or not there is a conspicuous peak in the middle frequency band (800 Hz to 2500 Hz), and can be determined based on the degree of concentration of energy in the middle frequency band.

Generally, the discriminating characteristic is a binary value of [+/-], that is, 0
Usually, it is expressed by the value of or 1. However, in the present invention, an intermediate value between 0 and 1 is used to improve the accuracy of word identification with respect to continuous changes in actual speech. FIG. 3 shows a correspondence between each of the above-described discriminating features and a phoneme that causes an increase in the discriminating feature amount.

Hereinafter, the method of calculating each of the discriminative feature amounts will be described in the fourth.
This will be described with reference to the flowcharts in FIGS.

FIG. 4A is a flowchart for obtaining the degree of the discriminative feature "strident". Hereinafter, a method of calculating the degree of the discriminative feature “strident” will be described in detail with reference to this flowchart.

In step S1, a discriminative feature amount “RSTR” representing the degree of the discriminative feature “strident” is calculated by the following equation.

Here, the numerator represents the energy amount weighting the high frequency band, and the discriminating feature amount “RSTR” represents the degree of energy concentration in the high frequency band. The above wi is a weight coefficient in each band, and its value is as shown in FIG. 4 (b). That is, the weight in the high frequency band is increased, and the concentration of energy in the high frequency band is determined efficiently. Α ₁ is a constant.

In step S2, it is determined whether or not the discriminative feature amount “RSTR” calculated in step S1 is a negative value. As a result, if the result is a negative value, the process proceeds to step S3, and if not, the operation of calculating the degree of the discriminative feature “strident” ends.

In step S3, the value of the discriminative feature “RSTR” calculated in step S1 is set to 0, and the discriminant feature “striden” is set.
The operation of calculating the degree of “t” ends.

FIG. 5 is a flowchart for obtaining the degree of the discriminative feature “voiced”. Hereinafter, the method of calculating the degree of the distinctive feature “voiced” will be described in detail with reference to this flowchart.

In step S21, the accumulated value of the feature amounts hi of the first to third channels is obtained, and it is determined whether or not the value is greater than the threshold value θv.

If the result is larger, the process proceeds to step S24, and if not, the process proceeds to step S22.

This is based on the fact that the distinctive feature "voiced" can be determined by the presence of energy in the frequency band from 120 Hz to 750 Hz.

In step S22, the ratio of the cumulative value of the feature value hi of the first to third channels calculated in step S21 to the cumulative value of the feature value hi in the entire frequency band is calculated, and whether the value is greater than the threshold θrv It is determined whether or not it is.

If the result is larger, the process proceeds to step S24, and if not, the process proceeds to step S23.

This is because the ratio of the cumulative value of the feature value hi to the cumulative value of the feature value hi in the entire frequency band is equal to or greater than θrv even if the cumulative value of the feature value hi of the first to third channels is determined to be smaller than the threshold value θv in step S21. If there is, the vocal cord vibration is determined to be present in the frequency bands of the first to third channels.

In step S23, assuming that there is no vocal fold vibration, the discriminating feature amount “RVOI” representing the degree of the discriminating feature “voiced” is set to 0, and the degree calculating operation of the discriminating feature amount “voiced” ends.

In step S24, assuming that the vocal cord vibration exists, the discriminative feature amount “RVOI” is set to 1, and the degree calculating operation of the discriminative feature “voiced” ends.

FIG. 6 is a flowchart for obtaining the degree of the discriminative feature “nasal”. Hereinafter, the method of calculating the degree of the distinctive feature “nasal” will be described in detail with reference to this flowchart.

 In step S31, 5 is set to "NCH".

In step S32, the amount “NAS1” of the spectrum unevenness in the low frequency band to the middle frequency band and the number of calculation “N1”
N "is calculated by the following equation.

In step S33, the value of the feature amount hi is equal to or larger than the threshold value θi.
The number of h. (that is, the spread of the spectrum from the low frequency band to the middle frequency band) “NBAND” is calculated by the following equation.

In step S34, it is determined whether “NAS1” is smaller than “RMAX”. If the result is smaller, the process proceeds to step S35; otherwise, the process proceeds to step S36.

Here, the initial value of "RMAX" is set to a large number that does not appear in the loop of this flowchart.

In step S35, the degree of spectrum flatness “RMAX” in the low frequency band to the middle frequency band is calculated by the following equation.

Here, the numerator indicates the spread of the low frequency band to the middle frequency band, and the denominator indicates the degree of unevenness in the low frequency band to the middle frequency band. Α ₃ is a constant.

 In step S36, 1 is added to "NCH".

In step S37, it is determined whether “NCH” is greater than seven. If the result is larger, the process proceeds to step S38; otherwise, the process returns to step S32 to further expand the calculation range and calculate “RMAX”.

In step S38, “NAS2” and “NAS3” are calculated by the following equations to calculate the degree of flatness of the spectrum in the frequency band corresponding to the first formant region.

In step S39, it is determined whether “NAS2” is larger than “NAS3”. If the result is larger, the process proceeds to step S40; otherwise, the process proceeds to step S41.

In step S40, the discriminating feature amount “RNAS” representing the degree of the discriminating feature “nasal” is calculated by the following equation, and the discriminating feature “nasal” degree calculating operation ends.

Here, the second term on the right side represents the degree of flatness of the spectrum in the first formant region. Α ₃ ′, α ₃ ″
Is a constant.

In step S41, the discriminating feature amount “RNAS” representing the degree of the discriminating feature “nasal” is calculated by the following equation, and the discriminating feature “nasal” degree calculating operation ends.

Here, the second term on the right side represents the degree of flatness of the spectrum in the first formant region.

FIG. 7A is a flowchart for obtaining the degree of the discriminative feature “murmur / buzz”. Hereinafter, a method of calculating the degree of the distinctive feature “murmur / buzz” will be described in detail with reference to this flowchart.

In step S51, a discriminative feature amount “RMUR” representing the degree of concentration of energy in the low frequency band is calculated by the following equation.

Here, wi is a weight coefficient in each band, and its value is as shown in FIG. 7 (b). That is, the weight coefficient in the low frequency band is increased to efficiently determine the concentration of energy in the low frequency band. Α ₄ is a constant.

In step S52, it is determined whether or not the discriminative feature amount “RMUR” calculated in step S51 is negative. If the result is negative, the process proceeds to step S53; otherwise, the discriminative feature “murmur / buzz” Ends the degree calculation operation.

In step S53, the discriminative feature amount “RMUR” calculated in step S51 is set to 0, and the discriminant feature “murmur /
The operation of calculating the degree of “buzz” ends.

FIG. 8 (a) is a flowchart for obtaining the degree of the discriminative feature "compact". Hereinafter, a method of calculating the discriminative feature “compact” will be described with reference to this flowchart.

 In step S61, "NCH" is set to 7.

In step S62, “NCOM” representing the energy amount in the middle frequency band is calculated by the following equation.

Here, wi is a weighting coefficient, and its value is as shown in FIG. 8 (b). In other words, "NC
The weight in the middle of the frequency band when calculating "OM" is increased so that the energy in the middle frequency band is efficiently obtained.

In step S63, it is determined whether or not the value of “NCOM” obtained in step S62 is greater than the value of “MAXC”. If the value is larger, the process proceeds to step S64; otherwise, the process proceeds to step S65.

Here, the initial value of the “MAXC” is set to a small number that does not appear in the loop of this flowchart.

In step S64, the value of "NCOM" obtained in step S62 is set in "MAXC".

 In step S65, 1 is added to the "NCH".

In step S66, “NCH” calculated in step S65 above
Is determined whether or not the value of is larger than 9. If the result is larger, the process proceeds to step S67; otherwise, the process returns to step S62 to calculate the next “NCOM” by further shifting the calculation range of “NCOM”.

In step S67, a discriminative feature amount “RCOM” representing the degree of concentration of energy in the middle frequency band is calculated by the following equation.

That is, the numerator corresponds to the energy amount in the middle frequency band, the denominator corresponds to the energy amount in the entire frequency band, and the value of the discriminative feature “RCOM” indicates the degree of concentration of the spectrum in the middle frequency band. I have.

FIG. 9 is a flowchart for obtaining the degree of the discriminative feature "diffuse". Hereinafter, the method of calculating the degree of the distinctive feature “diffuse” will be described in detail with reference to this flowchart.

In step S71, “NDF1” representing the amount of spectrum unevenness in the middle frequency band to the high frequency band is calculated by the following equation.

In step S72, the number of channels “NBAND” in which the value of the feature value hi is equal to or larger than the threshold θi in the middle frequency band to the high frequency band
The following equation is calculated. This “NBAND” is a value representing the spread of the spectrum in the middle frequency band to the high frequency band.

In step S73, the cumulative total “NDF2” of the feature amounts hi from the BPFs of the sixth to 16th channels and the cumulative total “NDF3” of the feature amounts hi from the BPFs of the third to 16th channels are calculated by equations. You.

In step S74, the discriminating feature amount “RDIF” representing the degree of flatness of the spectrum in the middle frequency band to the high frequency band is calculated by the following equation, and the degree calculating operation of the discriminating feature “diffuse” ends.

The first term on the right side of this equation represents the degree of flatness in the middle frequency band to the high frequency band, and the second term on the right side represents the degree of absence of the spectrum in the low frequency band to the middle frequency band. The threshold value θi is a threshold value provided for each BPF in order to obtain the spread of the spectrum, and α ₆ , α ₆ ′,
α ₆ ″ and α ₆ are constants.

FIG. 10 shows an example of an input voice waveform and a time-series pattern of each discriminative feature obtained from the input voice as described above. That is, FIG. 10 (a) shows the input voice waveform, and FIG.
FIG. 10 (b) is the phoneme label of the above speech waveform, FIG. 10 (c)
Is the discriminative feature “stride” in the input voice of FIG. 10 (a).
time series of the discriminative feature “RSTR” representing the degree of “nt”
FIG. 10D is a time series of the discriminative feature “RVOI” representing the degree of the discriminative feature “voiced”, and FIG. 10E is a discriminant feature “n”.
time series of the distinctive feature "RNAS", which represents the degree of "asal", 10th
FIG. 10 (f) is a time series of the discriminative feature “RMUR” representing the degree of the discriminant feature “murmur / buzz”, and FIG. 10 (g) is a discriminant feature “representing the degree of the discriminant feature“ compact ”. FIG. 10 (h) shows the time series of the discriminative feature “RDIF” representing the degree of the discriminative feature “diffuse”.

Next, the discriminative feature amount obtained for each phoneme in FIG. 10 will be discussed with reference to FIG. First, the distinctive feature “co
For the phonological / k / related to "mpact", the discriminative feature "RCO
Only the value of "M" is "1", and it can be said that the discriminative feature of the phoneme / k / is correctly extracted. The discriminative feature of the phoneme / a / related to the discriminative feature "voiced" is " RVOI ”is“ 1 ”and the value of the discriminative feature“ RCOM ”is“ 0 ”
A value of "1" indicates that the articulation position with the preceding phoneme / k / is a phoneme at the back, and the discriminative feature "RNAS" has a value of "0 to 1" in the latter half. This indicates that the phonation is affected by the following phoneme / n /, which means that the distinctive feature of the phoneme / a / is correctly extracted.
In the phonemes / n / related to "sal" and "murmur / buzz", the values of the discriminative features "RNAS" and "RMUR" are "1", and the discriminative features of the phoneme / n / are correctly extracted. In the phoneme / e / related to the discriminating feature "voiced", the value of the discriminating feature "RVOI" is "1", and the value of the discriminating feature "RNAS" is "0". It is a value of 1 ", indicating that it is nasalized by the effect of the previous phoneme / n /.
It can be said that the discriminative feature of e / is correctly extracted. Subsequently, in the phonemes / ts / related to the distinguishing features “strident” and “diffuse”, the distinguishing features “RSTR” and “R
The value of “DIF” is “1”, and it can be said that the discriminative feature of the phoneme / ts / is correctly extracted. The discriminant feature of the phoneme / u / related to the discriminant feature “voiced” is “RVOI”. It can be said that the value of "" is "1", and the discriminative feature amount of the phoneme / u / has been correctly extracted.

That is, it can be said that the discriminative feature amount extracted by the voice feature extraction method of the present invention is information that correctly represents the features of each phoneme in accordance with human articulation activity. Therefore, phonemes can be correctly identified based on the discriminating feature amount.

As described above, the voice feature extraction method of the present invention
The output power from the BPF is defined as a feature amount hi, and based on the feature amount hi, the discriminative feature amounts “RSTR”, “RVOI”, and “RNAS” indicating the degree of the discriminative feature which is a feature amount related to human articulation. ,
"RMUR", "RCOM" and "RDIF" are calculated. Therefore, according to the present invention, it is possible to extract information (discriminative feature amount) necessary for identifying each phoneme without being affected by a spectrum change caused by a speaker, and it is possible to extract the discriminative feature amount of an input voice. The similarity is obtained by DP matching or the like based on the discriminative feature amount of the standard pattern and the speech can be correctly recognized.

<Effects of the Invention> As is apparent from the above description, the speech feature extraction method according to the first aspect of the present invention uses a discriminative feature quantity representing frictional property, based on a spectral component of an input speech obtained by frequency analysis.
Discriminating features representing voiced, discriminating features representing flatness of the spectrum due to nasalization, discriminating features representing buzziness,
Since the discriminating feature indicating the flatness of the spectrum in the middle frequency band and the discriminating feature indicating the dispersion of the spectrum are obtained, the above-described six different discriminating features related to human articulation can be extracted. Therefore, according to the present invention, speech can be stably and correctly recognized without being affected by spectrum fluctuations caused by a speaker, and high recognition performance can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of a speech recognition device according to the present invention, FIG. 2 is a diagram showing frequency characteristics of each BPF,
FIG. 3 is a correspondence diagram between the discriminative feature and the phoneme which causes the increase of the discriminative feature, and FIG.
flowchart for determining the degree of "dent", FIG. 4 (b)
Is a diagram showing an example of a weighting factor, and FIG.
6 is a flowchart for determining the degree of the discriminative feature "nasal", FIG. 7A is a flowchart for determining the degree of the discriminative feature "murmur / buzz", and FIG. 7B ) Is a diagram showing an example of a weighting factor, FIG. 8 (a) is a flowchart for obtaining the degree of the discriminative feature “compact”, FIG. 8 (b) is a diagram showing an example of the weighting factor, and FIG. Fig. 10 is a flow chart for obtaining the degree of the feature "diffuse", and Fig. 10 is a diagram showing an example of a time-series pattern of a discriminative feature amount obtained based on the input speech waveform. ... Feature extraction unit, 4... Discriminative feature extraction unit, 5... Word recognition unit, 6.
... Standard pattern storage unit, 7... Result display unit.

Continuation of the front page (56) References JP-A-63-65500 (JP, A) JP-A-62-17900 (JP, A) JP-A-62-7081 (JP, A) JP-A-60-194500 (JP, A) JP-A-62-102298 (JP, A) JP-A-62-102297 (JP, A) JP-A-63-85697 (JP, A) JP-A-61-240300 (JP, A) 47-11635 (JP, B1) JP-B 47-11633 (JP, B1) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 7/08 G10L 9/00-9/20 JICST Science and Technology Reference file

Claims (1)

(57) [Claims] 1. A speech feature extraction method for frequency-analyzing an input speech and extracting speech features from the obtained spectrum components, wherein a frictional property is determined based on the spectrum components obtained by the frequency analysis. A first step of obtaining a discriminative feature amount representing the degree of the discriminative feature to be represented; and a second step of obtaining a discriminative feature amount representing the degree of the discriminative feature representing the voice based on the spectral component obtained by the frequency analysis. A second step, a third step of obtaining a discriminating feature amount indicating a degree of a discriminating feature indicating flatness of a spectrum by nasalization based on the spectrum component obtained by the frequency analysis, and A fourth step of obtaining a discriminating feature amount representing a degree of the discriminating feature representing buzziness based on the obtained spectral component; A fifth step of obtaining, based on the spectral components obtained by the analysis, a discriminating feature amount representing a degree of the discriminating feature representing the flatness of the spectrum of the middle frequency band; A sixth step of obtaining a discriminating feature amount representing a degree of a discriminating feature representing the dispersibility of the spectrum based on the speech, wherein a feature amount of a speech related to human articulation is extracted. Feature extraction method.