JP3129164B2 - Voice recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for recognizing a human voice by a machine.

[0002]

2. Description of the Related Art In recent years, a speech recognition apparatus for an unspecified speaker capable of recognizing anyone's voice without registering a user's voice has come into practical use. As a practical method for an unspecified speaker, a patent (JP-A-61-188599) will be described as a conventional example.

[0003] In the conventional method, a speech section is determined by finding the start and end of an input speech, and the speech section is linearly expanded and contracted by a predetermined time length (J frame), and the similarity between the speech section and a word standard pattern is statistically determined. This is a method of recognizing words obtained by performing pattern matching using a target distance scale.

Hereinafter, a conventional example will be described in detail with reference to FIGS. FIG. 7 is a flowchart showing the flow of processing of the conventional speech recognition method. In FIG. 7, 1 is an acoustic analysis unit, 2 is a feature parameter extraction unit, 3 is a voice section detection unit, 10 is a time axis linear normalization unit, 4 is a standard pattern storage unit, 11 is a distance calculation unit, and 6 is a distance comparison unit. It is.

In FIG. 7, when an input voice is input, the acoustic analysis unit 1 performs a linear prediction (LPC) analysis every analysis time (called a frame, 1 frame = 10 ms in the conventional example). Next, the feature parameter extracting unit 2 obtains P feature parameters for each frame. The characteristic parameter is L
PC mel cepstrum coefficient (9 in this example from C1 to C9)
), The normalized residual C0, and the time difference value V0 of the logarithmic power of the voice. Next, the voice section detector 3 detects the start frame and the end frame of the input voice. The method using voice power is the simplest method for detecting a voice section, but any method may be used. For the detected speech section, a time-series linear normalization unit 1 converts the feature parameter time series of the input speech.
At 0, linearly expands and contracts to the J frame. FIG. 8 conceptually illustrates this. Normally, J is set smaller than the actual number of frames of a word in order to reduce the amount of calculation and the number of estimated parameters of the standard pattern. This corresponds to thinning out frames at equal intervals over the entire speech section of a word. Assuming that the start frame of the detected input voice section is the first frame and the end frame is the I frame, the relationship between the j-th frame after expansion and contraction and the i-th frame of the input voice is

[0006]

(Equation 1)

[0007] Here, [] represents the largest integer not exceeding the number. The feature parameters for the J frames after expansion and contraction are arranged in time series to create an input time series pattern X.

[0008]

(Equation 2)

The distance between the input time-series pattern X and each standard pattern of the vocabulary to be recognized stored in the standard pattern storage unit 4 is determined by the distance calculation unit 11. A method of creating the standard pattern and a method of obtaining the distance will be described later. Finally, the distance comparing unit 6 has the smallest distance (similarity is maximum) among the distances from the respective standard patterns obtained by the distance calculating unit 11.
Is selected as a recognition result and output.

Hereinafter, a method of creating a word standard pattern and a method of calculating a distance between an input time-series pattern and a word standard pattern will be described.

The standard pattern of a certain word ωn is created by the following procedure. (1) Prepare M learning speech data in which many people (here, 100 people) uttered the word ωn. (2) Each data is linearly expanded and contracted using (Equation 1) and J
Normalize to frame. (3) The feature parameters after expansion and contraction are arranged in time series with respect to the m-th utterance data, and a time series pattern Cm is obtained. (m = 1, ..., M) (4) By using M time-series patterns Cm (m = 1, ..., M) to obtain their statistics (mean, covariance), Create a pattern.

This is obtained for each of the N words to be recognized. A time-series pattern Cm in which characteristic parameters after expansion and contraction are arranged in a time-series with respect to the m-th utterance data is expressed as follows.

[0013]

(Equation 3)

This is obtained for M learning speech data. By treating the time pattern Cm as one vector, the correlation between the parameter frames is taken into account. M J × P-dimensional vectors Cm (m = 1, ...,
From M), the average vector μ and the covariance matrix W are obtained. Hereinafter, the mean vector for the n-th word ωn is denoted by μn, and the covariance matrix is denoted by Wn.

The distance between the input time-series pattern X and the word standard pattern is calculated using a distance based on Bayesian judgment using a common covariance matrix.

The distance based on the Bayes judgment is obtained as follows. When the input vector X expressed by (Equation 2) is observed, the probability P (ωn | X) that it is the word ωn is obtained from Bayes' theorem.

[0017]

(Equation 4)

## EQU1 ## P (X | ωn) is the prior probability, the probability that the input vector X is observed when the input is the category ωn, and P (X) is the vector X when all possible inputs are considered. Probability. The appearance probability P (ωn) of the word ωn is assumed to be the same for each word and is assumed to be a constant. If the input X is assumed to be constant, P (X) becomes a constant. Therefore, the category ωn which maximizes the prior probability P (X | ωn) May be used as the determination result.

When the distribution of parameters is considered to be a normal distribution,
The prior probability P (X | ωn) is represented by (Equation 5).

[0020]

(Equation 5)

Here, t represents a transposed matrix. By taking the logarithm of both sides and omitting a constant term unnecessary for identification and further multiplying by -2, the following equation is obtained.

[0022]

(Equation 6)

This equation is a distance based on Bayes judgment for the word ωn. In order to reduce the amount of calculation and the number of estimated parameters, the covariance matrix is shared and this equation is developed into a linear linear discriminant. The covariance matrix Wn of each standard pattern of the vocabulary to be recognized is shared and is set to W. W is obtained by the following equation.

[0024]

(Equation 7)

Therefore,

[0026]

(Equation 8)

[0027] Substituting this into (Equation 6) and omitting the constant term unnecessary for identification,

[0028]

(Equation 9)

,

[0030]

(Equation 10)

[0031]

[Equation 11]

By the way,

[0033]

(Equation 12)

It can be seen that the linear primary discriminant is as follows. In this way, An and Bn are obtained for each vocabulary to be recognized, and stored in the standard pattern storage unit. The distance calculation unit uses the above equation to calculate the input time-series pattern X and the word ω
The distance Ln between n and the standard pattern is obtained.

[0035]

The conventional method is a practical method with a small amount of calculation. However, in the conventional method, the number J of frames of the standard pattern cannot be increased from the viewpoint of parameter estimation accuracy, and the entire voice section is recognized by thinning out frames at equal intervals.
For this reason, there is a problem that information of a portion such as a consonant, which has a short duration and needs to be compared in detail, is lost, and a sufficient speech recognition rate cannot be obtained. On the other hand, there is a problem that information in a portion that is temporally stationary and has a long continuous length like a vowel becomes redundant.

Further, the conventional method uses a statistical distance scale represented by a linear discriminant function as a single vector for the entire voice as a distance scale for matching between the input voice and the standard pattern. However, with the rapid increase in the speed of computers in recent years, it has become necessary to improve the recognition performance even if the amount of calculations increases.

Furthermore, the conventional method requires learning speech data uttered by a large number of people in order to create a word standard pattern, so that it is not easy to change the vocabulary to be recognized.

The present invention solves the above-mentioned conventional problems, and the first object of the present invention is to provide a speech recognition method that improves the recognition rate as compared with the conventional example.

A second object is to provide a speech recognition method that further improves the recognition rate by using a distance scale having high discrimination performance.

A third object of the present invention is to provide a high-accuracy speech recognition method capable of easily changing a vocabulary to be recognized, capable of creating a word standard pattern from Japanese kana character notation.

[0041]

Means for Solving the Problems First, the present invention has solved the above-mentioned problems by the following means.

The consonant part in the word voice creates a standard pattern by using a series of frames around the reference frame, and the vowel part creates a standard pattern by linearly expanding and contracting the frame. At the time of recognition, the consonant part performs collation of frames continuously, and the vowel part performs collation by expanding and contracting the frame. Speech recognition performance can be improved by taking such a frame.

In order not to increase the amount of calculation and the number of estimated parameters of the standard pattern, the collation between the input speech and the standard pattern is performed by using a statistical distance represented by a linear discriminant function considering the inter-frame correlation with the entire speech as one vector. Use a scale. Alternatively, although the amount of calculation is doubled, frames are treated independently, and instead, a statistical distance scale represented by a linear discriminant function using a dynamic feature parameter that is a time variation of a feature parameter is used.

Secondly, the present invention has solved the above problem by the following means. A statistical distance scale represented by a quadratic discriminant function is used as a distance scale for matching the input voice with the standard pattern in the first means. However, when trying to create a standard pattern using the time-series pattern of the whole word of the feature parameter as one vector, a huge number of training samples are required for estimating the covariance, so the time pattern is set as an independent vector for each frame. deal with. By using a statistical distance measure represented by a quadratic discriminant function, the speech recognition performance can be further improved. When a dynamic feature parameter that is a time change amount of the feature parameter is used together, the speech recognition performance can be further improved.

Third, the present invention has solved the above problem by the following means. Syllable, CV (consonant + vowel), VC
(Vowel + consonant), VCV (vowel + consonant + vowel), or C
A standard pattern is created for each unit such as VC (consonant + vowel + consonant) in the same manner as the first and second means, and these are connected to create an arbitrary word standard pattern. Recognize in the same way as the means. Because it is possible to create word standard patterns according to Japanese kana character notation,
The vocabulary to be recognized can be easily changed.

[0046]

[Function] Japanese is composed of consonants and vowels. In general, the vowel part has a feature that the spectrum of the vowel part is stationary with little temporal change, and its continuation length easily expands and contracts due to a difference in utterance speed. On the other hand, the consonant part has information for identifying a phoneme in the temporal change of the spectrum, and has a feature that its continuation length is relatively short and does not easily expand or contract even if the utterance speed is different.

In the present invention, first, the consonant part is obtained by successively taking frames around the reference frame and performing collation without expanding / contracting, and the vowel part is collated by expanding / contracting the frame, whereby local consonant parts are collated. The characteristics of the temporal change of the spectrum and the characteristics of the global spectrum of the vowel part can be appropriately captured without being affected by the utterance speed, and the recognition performance is improved. By reducing the number of vowel frames instead of taking the consonant portions of the standard pattern continuously, the number of frames of the standard pattern does not increase.

Using a statistical distance scale represented by a first-order discriminant function taking the inter-frame correlation into consideration assuming that the entire speech is one vector, the recognition rate can be improved without increasing the amount of calculation and the number of estimated parameters. it can. When a frame is treated independently and a dynamic distance parameter, which is a time variation of a characteristic parameter, is used in combination and a statistical distance scale represented by a linear discriminant function is used, the calculation amount is doubled.
The recognition rate can be improved.

Second, the present invention can further improve speech recognition performance by treating frames independently and using a statistical distance scale represented by a quadratic discriminant function when matching an input speech with a standard pattern. it can. When the dynamic feature parameter, which is the amount of time change of the feature parameter, is used together, the feature amount of the time change lost by treating the frame independently can be captured, so that the speech recognition performance can be further improved. it can.

Third, the present invention connects a standard pattern such as syllable, CV (consonant + vowel), VC (vowel + consonant), VCV (vowel + consonant + vowel) or CVC (consonant + vowel + consonant). By creating and recognizing an arbitrary word standard pattern, a word standard pattern can be created in accordance with Japanese kana character notation, so that the vocabulary to be recognized can be easily changed.

Further, by introducing the word spotting function, a highly practical recognition device that is robust against noise can be realized.

[0052]

【Example】

Embodiment 1 Hereinafter, a first embodiment of the present invention will be described.

In the first embodiment, a single syllable uttered independently of a syllable, which is the minimum unit of Japanese utterance, is used as a recognition target, and Bayes judgment is performed in which the entire utterance is a single vector and the covariance matrix is shared. A speech recognition method will be described in which an input speech is compared with a single syllable standard pattern using a statistical distance scale represented by a primary discriminant function based on the standard distance function.

In the first embodiment, a single syllable section is detected by detecting a single syllable section of the unknown input voice and comparing it with a single syllable standard pattern created in advance.

A Japanese monosyllable is composed of a consonant part and a vowel part following it. In general, the vowel part has a feature that the spectrum of the vowel part is stationary with little temporal change, and its continuation length easily expands and contracts due to a difference in utterance speed. On the other hand, the consonant part has information for identifying a phoneme in the temporal change of the spectrum, and has a feature that its continuation length is relatively short and does not easily expand or contract even if the utterance speed is different. Therefore, the consonant part continuously compares frames (analysis time unit; one frame = 10 ms in this embodiment) and compares the input voice with the standard pattern without expanding and contracting, and the vowel part performs expansion and contraction of the frame. Since the vowel part has a steady spectrum, even if a few frames adjacent to each other are grouped together and a standard pattern of one frame is used, a decrease in the discrimination performance is small. By reducing the number of frames in the vowel part instead of the number of frames in the consonant part, the recognition rate can be improved without increasing the number of frames in the entire single syllable standard pattern.

The first embodiment will be described with reference to FIGS. 1, 2 and 3. FIG. 1 is a flowchart showing the flow of processing of the voice recognition method according to the first embodiment. In FIG. 1, reference numeral 1 denotes an acoustic analysis unit that performs linear prediction (LPC) analysis of an unknown input voice for each analysis time (frame), 2 denotes a feature parameter extraction unit that obtains a feature parameter for each frame, and 3 denotes a start frame of the input voice and A voice section detection unit that detects the end frame, a standard pattern storage unit that stores a single syllable standard pattern, a DP matching unit that calculates the distance between the input voice and the single syllable standard pattern, and a DP matching unit that calculates a distance between the input speech and the single syllable standard pattern. A distance comparison unit that recognizes a speech name corresponding to the standard pattern having the smallest value (similarity is maximum) among the distances from the respective standard patterns.

Next, the operation will be described. The monosyllable standard pattern is created in advance and stored in the standard pattern storage unit 4. A method for creating a single syllable standard pattern will be described later.
When an unknown input voice is input, the acoustic analysis unit 1 performs a linear prediction (LPC) analysis for each analysis time (frame).
Next, P (P is a positive integer) in the feature parameter extraction unit 2
Is obtained for each frame. The feature parameters are LPC mel-cepstral coefficients (in this example, C1 to C9
9), the normalized residual C0, and the time difference value V0 of the logarithmic power of the voice. Next, the voice section detection unit 3 detects the start frame and the end frame of the input voice using voice power information and the like. In the first embodiment, voice section detection uses voice power, but any method may be used. Next, in the DP matching unit 5, the feature parameter time series of the input voice and a certain single syllable standard pattern stored in the standard pattern storage unit 4 are dynamically checked by the DP method, and the single syllable standard pattern is compared. Find the distance. This is determined for all monosyllables to be recognized. DP
The method of collation and distance calculation will be described later. Finally, the distance comparison unit 6 selects, as a recognition result, a speech name corresponding to the standard pattern having the minimum value (similarity is maximum) among the distances from the respective standard patterns obtained by the DP matching unit 5, Output.

Hereinafter, a method of creating a single syllable standard pattern will be described. A speech standard pattern for speaker-independent speech recognition is created by obtaining the statistics (average value, covariance) using learning speech data uttered by many people.

A Japanese monosyllable is composed of a consonant part followed by a vowel part. The single-syllable standard pattern is created from a set of vectors in which frames are extracted non-linearly from the learning voice data of the same category (single syllable), and the feature parameters of these frames are arranged in time series. The method of extracting a frame in a non-linear manner is as follows.

A consonant has information for identifying a phoneme in a temporal change of a spectrum, and has a feature that its continuation length is relatively short and does not easily expand or contract even if the utterance speed is different. Therefore, for the consonant part, a temporal position that best represents the feature of the consonant is set as a reference frame, and several frames before and after each reference frame are continuously extracted from the learning voice data. The vowel part is extracted by expanding and contracting the frame linearly from the end of the continuous time pattern to the end frame of the voice. FIG. 2 shows a conceptual diagram thereof.

In FIG. 2, the reference frame of the consonant is visually labeled as a phoneme label 21 on the learning speech data based on a certain standard determined for each consonant. In the present embodiment, the unvoiced plosive (/ c /, / p /, /
t /, / k /) are plosive frames, nasal sounds (/ m /, / n /) and unvoiced fricatives (/ h /, / s /) are vowels, voiced plosives (/ b /,
/ d /, / g /, / r /) is the burst frame (end of buzz bar), / z /
Designates a portion from voiced to unvoiced as a reference frame. In addition, single vowels (“A”, “I”, “U”,
"E", "O") and semi-vowels ("Ya", "Yu", "Yo",
“Wa”) defines the rising frame of the voice power 22 at the beginning of the word as a reference frame. Then, in the characteristic parameter time series 23, the L1
The frame and the subsequent L2 frame are continuously extracted. L1 and
The value of L2 differs for each consonant. L1 and L2 were determined by examining the valid frames for discriminating consonants by preliminary experiments. Further, the vowel part from the end frame of the continuous time pattern to the end frame of the syllable is linearly expanded and contracted and extracted, thereby creating the time-series pattern Cm24. The / j / of the resonate is characterized by a slow spectral transition from the consonant to the succeeding vowel, and tends to expand and contract according to the utterance speed.

The standard pattern of a single syllable ωn is created in the following procedure. (1) M that many people (here 100 people) uttered a monosyllable ωn
The learning voice data is prepared. (2) Each data is nonlinearly expanded and contracted and normalized to a J frame. (3) The feature parameters after expansion and contraction are arranged in time series with respect to the m-th utterance data, and a time series pattern Cm is obtained. (m =
(4) A standard pattern is created by using the M time-series patterns Cm (m = 1,..., M) to find their statistics (mean, covariance). I do.

A method for obtaining the time-series pattern Cm from the m-th learning voice data will be described.

It is assumed that the number of frames of the standard pattern is J frames, and L frames (L = L1 + L2 + 1) among these are continuous. ｛Reference frame-L of the m-th learning voice data
Assuming that the 11 frame is the first frame and the end frame of the voice section is the I frame, the relationship between the i-th frame of this data and the j-th frame after expansion / contraction is represented by (Expression 13). Here, [] represents the largest integer not exceeding the number. First
In this embodiment, J = 20 and L = 10. J must be the same value for all monosyllables, but L may be different for each monosyllable.

[0065]

(Equation 13)

The feature parameters for the J frames after expansion and contraction are arranged in time series to create a time pattern Cm.

[0067]

[Equation 14]

This is obtained for M pieces of learning speech data. By treating the time pattern Cm as one vector, the correlation between the parameter frames is taken into account. M J × P-dimensional vectors Cm (m = 1, ...,
From M), the average vector μ and the covariance matrix W are obtained.

Further, this is obtained for each of the N single syllables to be recognized. Hereinafter, the average value vector for the n-th single syllable ωn is denoted by μn, and the covariance matrix is denoted by Wn.

The distance between the time-series pattern of the characteristic parameters of the unknown input speech and the standard single-syllable pattern is calculated using a distance based on Bayesian judgment using a common covariance matrix.

The distance based on the Bayes judgment is obtained as follows. Now, an input vector X that can be obtained by arranging the feature parameters of the unknown input voice after expansion and contraction for J frames is

[0072]

(Equation 15)

When the input vector X is observed, the probability P (ωn | X) that it is a single syllable ωn is obtained in the same manner as in the conventional example. From Bayes' theorem, P (ωn | X) is

[0074]

(Equation 16)

Is obtained. P (X | ωn) is the prior probability, the probability that vector X is observed when the input is category ωn, and P (X) is the vector X when all possible inputs are considered. Probability. The appearance probability P (ωn) of the word ωn is assumed to be the same for each word, and is assumed to be a constant.
Is constant, P (X) is a constant, so the prior probability P
The category ωn that maximizes (X | ωn) may be used as the determination result.

Assuming that the parameter distribution is a normal distribution,
The prior probability P (X | ωn) is represented by (Equation 17).

[0077]

[Equation 17]

Here, t represents a transposed matrix. By taking the logarithm of both sides and omitting a constant term unnecessary for identification and further multiplying by -2, the following equation is obtained.

[0079]

(Equation 18)

This equation is a distance based on Bayes judgment for a single syllable ωn. Here, in order to reduce the amount of calculation and the number of estimated parameters, the covariance matrix is shared as in the conventional example, and this equation is developed into a linear discriminant. The covariance matrix Wn of each single syllable standard pattern is shared, and is set to W. W is obtained by the following equation.

[0081]

[Equation 19]

Therefore,

[0083]

(Equation 20)

[0111] Substituting this into (Equation 18) and omitting the constant term unnecessary for identification

[0085]

(Equation 21)

[0086]

[0087]

(Equation 22)

[0088]

(Equation 23)

By setting

[0090]

(Equation 24)

It can be seen that the linear primary discriminant is as follows. In this way, An and Bn are obtained for each single syllable to be recognized and stored in the standard pattern storage unit 4.

Hereinafter, a detailed description will be given of a method in which the DP collation unit 5 performs dynamic time matching between the input speech and the single syllable standard pattern by the DP method to collate and obtain a distance.

The start frame of the voice section detected by the voice section detection unit is the first frame, and the end frame is the I-th frame. Xi is a sequence of P feature parameters of the i-th frame of the input voice.

[0094]

(Equation 25)

It is assumed that Then, r (1), r (2),
.., R (j),..., R (J) -th frame x are arranged to create a time pattern X for J frames. This becomes the input vector.

[0096]

(Equation 26)

The standard patterns of a single syllable ωn are An and Bn, and An is

[0098]

[Equation 27]

When writing, the input vector X and the monosyllable ωn
The distance Ln from the standard pattern is

[0100]

[Equation 28]

Therefore,

[0102]

(Equation 29)

Is obtained. Therefore, r that minimizes Ln
(j) may be obtained by the DP method. The value at which Ln is minimum is obtained by the DP method using the following recurrence formula.

[0104]

[Equation 30]

Where m is an integer from ms to me and ms, m
The value of e differs for each single syllable and for each frame of the standard pattern. In the continuous part from j = 1 to j = L

[0106]

(Equation 31)

Then, the input voice is collated continuously with the standard pattern without expanding / contracting. In this embodiment, the values of ms and me of the expansion and contraction part are the standard patterns of the single syllable.

[0108]

(Equation 32)

It was decided to expand and contract between the two. These DP paths are shown in FIG. 3A for the continuous portion, and FIG. 3B for the stretchable portion.

The distance Ln between the input vector X and the standard pattern of the single syllable ωn is obtained by subtracting the cumulative distance g (I, J) of the last frame of the single syllable standard pattern in the last frame of the input speech from Bn.

[0111]

[Equation 33]

This is obtained for all single syllable standard patterns. In the first embodiment, the method of performing the matching after detecting the voice section of the input voice has been described.
Without detecting the voice section of the input voice, for all the input voice sections including noise,

[0113]

(Equation 34)

When continuous DP matching is performed by a recurrence formula expressed by the following formula, an input frame i with which g (i, J) is minimized is obtained.

[0115]

(Equation 35)

By setting the distance from the standard pattern of a single syllable ωn, recognition can be performed without detecting a voice section. This is called word spotting.

However, when performing word spotting, it is necessary to use a posteriorized distance scale. The method is as follows. In (Equation 16), when word spotting is performed, the input X in different input sections must be compared, so that the input X is not constant. Therefore, the category ωn that maximizes the posterior probability P (ωn | X) in consideration of the term of P (X) needs to be set as the determination result.

P (X) is the probability of observing the vector X when all possible inputs are considered. Therefore, as a surrounding information pattern for posterior stochasticization, an average value vector and a covariance matrix of all possible inputs are obtained. That is, the average value vector μe is obtained from the J-frame time-series pattern created by shifting the J-frame time window by one frame with respect to the feature parameter time-series of all monosyllable learning speech data to be recognized.
And a covariance matrix We are obtained in advance. However, in order to spot an uttered voice from a section containing noise, it is necessary to create a surrounding information pattern including a noise section for posterior probability. P (X) is obtained from the average value vector μe of the surrounding information pattern and the covariance matrix We.

Assuming that the parameter distribution is a normal distribution,
The posterior probability P (ωn | X) is represented by (Equation 36).

[0120]

[Equation 36]

Here, t represents a transposed matrix. Taking the logarithm of both sides and multiplying by -2 gives the following equation.

[0122]

(37)

This equation is a distance based on Bayesian judgment made into a posteriori probability for a single syllable ωn. Here, in order to reduce the amount of calculation and the number of estimated parameters, the covariance matrix is shared and this equation is developed into a linear discriminant. The covariance matrix Wn of the standard pattern of each vocabulary to be recognized and the covariance matrix We of the surrounding information pattern are shared, and are set to W. W is obtained by the following equation. g is the ratio of mixing surrounding information patterns, and here, g = N.

[0124]

(38)

Therefore,

[0126]

[Equation 39]

[0127] Substituting this into (Equation 37) gives

[0128]

(Equation 40)

Becomes

[0130]

[Equation 41]

[0131]

(Equation 42)

By setting

[0133]

[Equation 43]

It can be seen that the linear primary discriminant is as follows. When word spotting is performed, An and Bn are obtained for each of the single syllables to be recognized in this way, and stored in the standard pattern storage unit 4.

Note that, for a phoneme such as an unvoiced fricative or a buzz bar at the beginning of a word, which has a steady spectrum and expands and contracts sharply due to vocalization, a similar part to a vowel part is temporally preceding a continuous pattern centered on a reference frame. A pattern that linearly expands and contracts may be provided.

In the first embodiment, an example in which a single syllable is recognized has been described, but word recognition can be performed in a similar manner. In this case as well, the standard pattern is such that the consonant part is continuously expanded around the reference frame, and the vowel part is expanded and contracted linearly, so that the J
Create to be a frame. When recognizing, the collation is performed by the DP method in the same manner as in the first embodiment, while keeping the continuous portion from expanding and contracting.

(Embodiment 2) Hereinafter, a second embodiment of the present invention will be described.

In the second embodiment, Japanese monosyllables are to be recognized, and a statistical distance scale represented by a secondary discriminant function based on Bayesian judgment is used for each frame of the input speech and monosyllable standard patterns. A speech recognition method will be described in which the obtained feature parameter vector and the dynamic feature parameter vector are collated and recognized.

In the second embodiment, a single syllable is recognized by detecting a single syllable section of the unknown input voice and comparing it with a previously prepared single syllable standard pattern as in the first embodiment. Perform

A second embodiment will be described with reference to FIG. FIG. 4 is a flowchart showing the flow of the process of the second embodiment.

In FIG. 4, reference numeral 1 denotes an acoustic analysis unit for performing LPC analysis on an unknown input voice for each frame, 2 denotes a feature parameter extraction unit for obtaining feature parameters for each frame, and 7
Is a dynamic feature parameter extraction unit for calculating the time variation of feature parameters, 3 is a speech section detection unit for detecting the start and end frames of the input speech, 4 is a standard pattern storage unit for storing a single syllable standard pattern, and 5 is an input speech. A DP comparing unit 6 for obtaining a distance between the reference pattern and the single syllable standard pattern, and a distance comparing unit 6 for recognizing a speech name corresponding to the standard pattern having the smallest value among the distances obtained by the DP matching unit 5.

Next, the operation will be described. The monosyllable standard pattern is created in advance and stored in the standard pattern storage unit 4. A method for creating a single syllable standard pattern will be described later.
When an unknown input voice is input, the acoustic analysis unit 1 performs LPC analysis for each frame, and the feature parameter extraction unit 2 performs PPC analysis.
The number of feature parameters is determined for each frame. The characteristic parameters are the same as in the first embodiment. Then, for each dimension of the feature parameter, the dynamic feature parameter extraction unit 7 obtains P regression coefficients, which are time variation amounts, for each frame. Next, the voice section detection unit 3 detects the start and end frames of the input voice, and the DP matching unit 5 calculates a statistical distance represented by a quadratic discriminant function between the feature parameter time series of the input voice and the single syllable standard pattern. Matching is dynamically performed by the DP method using the scale, and the distance to each monosyllable standard pattern is obtained.
Finally, the distance comparison unit 6 selects and outputs, as a recognition result, a speech name corresponding to the standard pattern having the minimum value among the distances from the respective standard patterns obtained by the DP matching unit 5.

The distance between the time-series pattern of the characteristic parameter of the unknown input speech and the standard single-syllable pattern is calculated using a distance based on Bayesian judgment.

The distance based on the Bayes judgment is a quadratic discriminant function, and the amount of calculation is proportional to the square of the number of dimensions of the vector for which the distance is to be obtained. Therefore, when the number of dimensions of the vector is large, the amount of calculation explosively increases. In addition, a large number of training samples are required for estimating the covariance. Therefore, it is necessary to reduce the number of dimensions of the vector. In the first embodiment, the distance between the input speech and the single syllable standard pattern is obtained using the time-series pattern of the entire single syllable of the feature parameter as one vector. In the second embodiment, the distance is divided for each frame. deal with. That is, a P-dimensional vector composed of P feature parameters is represented by J
The arrangement of the frames is used as a standard pattern, the distance between each frame and the corresponding frame of the input voice is determined by the distance based on Bayesian judgment, and the sum is used as the distance between the input voice and the single-syllable standard pattern. However, if the frames are handled independently in this way, it is not possible to capture dynamic changes in the feature parameters. Therefore, the time variation of the feature parameter is introduced as a dynamic feature parameter. In this embodiment, the regression coefficients of the p-th feature parameter of two frames before and after a certain frame (a total of five frames) are set as the p-th dynamic feature parameters of the frame. The dynamic feature parameter extraction unit 7 obtains P dynamic feature parameters for each frame.

Now, a vector composed of P feature parameters of the i-th frame of the unknown input speech is

[0146]

[Equation 44]

Further, a vector composed of P dynamic feature parameters is represented by

[0148]

[Equation 45]

It is assumed that In the same manner as in the first embodiment, the monosyllabic standard pattern expands and contracts each of the learning speech data nonlinearly and normalizes it to a J frame, and averages the characteristic parameters of the jth frame with respect to the nth monosyllable ωn. Vector μnj, covariance matrix Wnj, mean vector of dynamic feature parameters

[0150]

[Outside 1]

And the covariance matrix

[0152]

[Outside 2]

Are obtained for J frames from j = 1 to J, and these are stored in the standard pattern storage unit 4.

At this time, the distance based on the Bayes judgment between the input i-th frame and the j-th frame of the single syllable ωn is expressed by (Equation 46).

[0155]

[Equation 46]

Here, t represents a transposed matrix. 1,2,…, j,…, J-th frame of the standard pattern for monosyllable ωn,
When the r (1), r (2),..., R (j),..., R (J) -th frames of the input voice correspond to each other, the distance Ln between the input voice and the monosyllable ωn is

[0157]

[Equation 47]

It is assumed that Therefore, (Equation 46) (Equation 47)
Than

[0159]

[Equation 48]

Is as follows. Therefore, r such that Ln is minimized
(j) may be obtained by the DP method. The value at the time when Ln becomes the minimum is obtained by the following recurrence formula by the DP method as in the first embodiment.

[0161]

[Equation 49]

Here, m is an integer from ms to me, and ms and m
The value of e is the same as in the first embodiment. In the continuous part (Equation 3
1) and the collation is performed without expanding / contracting.

The cumulative distance g (I, J) of the last frame of the single syllable standard pattern in the end frame of the input voice is the distance Ln between the input voice and the single syllable ωn standard pattern.

[0164]

[Equation 50]

This is obtained for all monosyllable standard patterns. In the second embodiment, since the distance is calculated independently for each frame, the number of frames of the standard pattern is
It may be different for each single syllable. In that case, the distance Ln between the input voice and the monosyllable ωn is

[0166]

(Equation 51)

It is assumed that Here, Jn is the number of frames of a single syllable ωn. In the second embodiment, since the distance based on the Bayes determination is used, the calculation amount is larger than that of the conventional example. In the conventional example and the first embodiment, since the distance based on the Bayes decision in which the covariance matrix is shared by using the entire speech as one vector is used, if the number of frames is J and the number of parameters per frame is P, 1 The number of calculations of the sum of products per syllable is JP times. This is 220 times if J = 20 and P = 11. On the other hand, in the distance based on Bayesian judgment, if the number of dimensions of the vector is P, the number of times of product sum calculation is P (P + 3) / 2
Times. When a frame is treated independently and a feature parameter vector and a dynamic feature parameter vector are used, the number of product sum calculations per frame is P (P + 3) / 2 × 2
Therefore, JP (P + 3) times in the J frame. This is 3080 times if J = 20 and P = 11. That is, the product-sum calculation amount of the second embodiment is 14 times that of the conventional example.

In the second embodiment, a statistical distance scale represented by a quadratic discriminant function based on Bayes decision is used as a distance measure for collation. However, a primary measure based on Bayes decision using a common covariance matrix is used. A statistical distance measure represented by a discriminant function can also be used. As a result, the amount of calculation is about twice that of the conventional example, and a higher recognition rate than the conventional example can be obtained.

In the second embodiment, the feature parameter vector obtained for each frame of the input speech and the single syllable standard pattern is compared with the dynamic feature parameter vector for recognition. However, the feature parameter vector may be used alone. Good. In this case, the recognition rate is slightly lowered, but there is an advantage that the calculation amount is reduced by half.

Also, by performing continuous DP matching as in the first embodiment, word spotting can be performed. When word spotting is performed, it is necessary to use a posterior-probabilistic distance measure as a distance measure in order to compare different input sections. The method is as follows.

As the surrounding information pattern for posterior stochasticization, it is necessary to obtain an average value vector and a covariance matrix for all possible inputs. An average vector μ of feature parameters of one frame created for all speech sections of all syllable learning speech data to be recognized.
e, covariance matrix We, mean vector of dynamic feature parameters

[0172]

[Outside 3]

And the covariance matrix

[0174]

[Outside 4]

Are obtained and stored in the standard pattern storage unit 4 as standard patterns. However, in order to spot an uttered voice from a section containing noise, it is necessary to create a surrounding information pattern including a noise section for posterior probability.

The distance based on the Bayesian judgment made into the posterior probability is expressed by (Equation 52).

[0177]

(Equation 52)

Therefore, the distance Ln between the input voice and the monosyllable ωn is calculated by using (Equation 53) instead of (Equation 48), and
Uses (Expression 54) instead of (Expression 49).

[0179]

(Equation 53)

[0180]

(Equation 54)

Embodiment 3 Hereinafter, a third embodiment of the present invention will be described.

In the third embodiment, a syllable is cut out from the learning word voice data, and a syllable standard pattern is created from the feature parameter vector and the dynamic feature parameter vector for each frame in the same manner as in the second embodiment. A method of recognizing a word by creating a word standard pattern by concatenating them and performing collation using a statistical distance scale represented by a secondary discriminant function based on Bayes judgment in the same manner as in the second embodiment. explain.

A third embodiment will be described with reference to FIGS. FIG. 5 is a flowchart showing the flow of the process of the third embodiment.

In FIG. 5, reference numeral 1 denotes an acoustic analysis unit for performing LPC analysis of an unknown input speech for each frame, 2 a feature parameter extraction unit for obtaining a feature parameter for each frame, and 7 a dynamic feature parameter for obtaining a time variation of the feature parameter. Extraction unit, 3 is a voice section detection unit that detects the start and end frames of the input voice, 8 is a kana notation word dictionary, 9 is a syllable standard pattern storage unit that stores syllable standard patterns, and 5 is the input voice and each word standard pattern. Matching unit for calculating distance of the object, 6
Reference numeral denotes a distance comparison unit that recognizes a speech name corresponding to a standard pattern having the smallest value (similarity is maximum) among the distances obtained by the DP comparison unit 5.

Next, the operation will be described. The syllable standard pattern is created in advance and stored in the syllable standard pattern storage 9. A method for creating the syllable standard pattern will be described later.
When an unknown input voice is input, the acoustic analysis unit 1 performs LPC analysis for each frame, and the feature parameter extraction unit 2 performs PPC analysis.
The number of feature parameters is determined for each frame. The characteristic parameters are the same as in the first embodiment. Then, for each dimension of the feature parameter, the dynamic feature parameter extraction unit 7 obtains P regression coefficients, which are time variation amounts, for each frame. Next, the voice section detector 3 detects the start and end frames of the input voice. Next, the syllable standard patterns stored in the syllable standard pattern storage unit 9 are linked according to the kana character notation of the words written in the kana notation word dictionary 8 to create a word standard pattern. As in the second embodiment, the DP collating unit 5 dynamically collates the feature parameter time series of the input speech and each word standard pattern by the DP method, and obtains a distance to each word standard pattern. Finally, the distance comparison unit 6 selects, as a recognition result, a speech name corresponding to the standard pattern having the minimum value (similarity is maximum) among the distances from the respective standard patterns obtained by the DP matching unit 5, Output.

Hereinafter, a method for creating a syllable standard pattern will be described with reference to FIG. In consideration of the phonemic environment, voice data in which various people uttered various word sets with balanced phonemes are prepared as learning voice data. In the learning speech data, the start and end positions of the syllable 64 and the reference frame of the consonant are visually checked beforehand for the phoneme label 6.
Labeling is performed as 1. Then, speech data from the beginning to the end of each syllable is cut out, and for each syllable, the consonant part is continuously expanded and contracted linearly around the reference frame as in the second embodiment, and the characteristic parameters of the syllable standard pattern are obtained. A time series 63 is created. For phonemes whose spectrum is steady and sharply expands and contracts due to utterance, such as unvoiced fricatives and buzz bars at the beginning of the word, a linearly expanding and contracting pattern similar to the vowel part is added to the temporally preceding part of the continuous pattern centered on the reference frame. It may be provided.

When the input speech is collated with the word standard pattern by time expansion and contraction by the DP method, as in the second embodiment, the consonant portion is collated from the beginning to the end of the word without being expanded or contracted. Perform The DP path may be changed for each frame so as to reach the range represented by (Equation 32) for each syllable, or the length of the syllable standard pattern may be changed for each syllable such as 1/2 of the average continuous length of the syllable. If it changes to, it may be uniform with the expansion and contraction part.

In the third embodiment, recognition is performed in syllable units, but CV (consonant + vowel), VC (vowel + consonant), V
A speech piece such as CV (vowel + consonant + vowel) or CVC (consonant + vowel + consonant) may be used as a unit. Also in this case, the consonant part continuously performs collation centering on the reference frame. FIG. 6B is an example of a cutting method when the recognition unit is CV / VC.

In the third embodiment, a statistical distance scale represented by a quadratic discriminant function based on Bayesian judgment is used as a distance scale for collation. However, a primary scale based on Bayesian judgment using a common covariance matrix is used. A statistical distance measure represented by a discriminant function can also be used. This makes it possible to realize a speech recognition method that can easily change the recognition target vocabulary with a small amount of calculation.

In the third embodiment, the feature parameter vector obtained for each frame of the input speech and the monosyllable standard pattern is compared with the dynamic feature parameter vector for recognition. However, the feature parameter vector may be used alone. Good. In this case, the recognition rate is slightly lowered, but there is an advantage that the calculation amount is reduced by half.

By performing continuous DP matching, word spotting can be performed in the same manner as in the first and second embodiments.

In order to confirm the effects of the first, second, and third embodiments, a recognition experiment was performed using 110 monosyllabic voices uttered by a total of 150 men and women and 100 word voices of place names.
A voice standard pattern was created using the data of 100 people (50 men and women), and the data of the remaining 50 people were evaluated.

Table 1 shows the evaluation conditions. (Table 2) shows the recognition rate of 110 single syllables and 100 words of place names according to the conventional example.
The recognition rate of 110 single syllables according to the first embodiment, the recognition rate of 110 single syllables according to the second embodiment, and the recognition rate of 100 place names according to the third embodiment are shown.

[0194]

[Table 1]

[0195]

[Table 2]

In Table 2, the amount of calculation is the sum of products when calculating the distance between the input voice and the standard pattern when the number J of frames of the standard pattern is 20 and the number P of feature parameters per frame is 11. The ratio of the number of operations when the method according to the conventional example is set to 1 is shown. In the method according to the third embodiment, the distance is calculated only for the total number of frames of syllables appearing in 100 words of the place name, so that the calculation amount does not increase so much.

As described above, in the method according to the first embodiment,
Single syllable recognition rate is 68.0% compared to 47.2% of the conventional method
Thus, the recognition rate can be improved without increasing the amount of calculation and the number of estimated parameters.

In the method according to the second embodiment, the monosyllable recognition rate is 75.4% as compared with the method according to the first embodiment.
The recognition rate can be further improved.

In the conventional method, it was difficult to change the vocabulary to be recognized. However, in the method according to the third embodiment, a word standard pattern can be created from kana notation. The word recognition rate also improved from 97.3% of the conventional method to 98.9%.

Each of the embodiments is a method capable of word spotting. By introducing word spotting, a highly practical recognition device that is robust against noise can be realized.

[0201]

According to the present invention, first, the consonant part continuously takes a frame around the reference frame, and the vowel part is linearly expanded and contracted to create a standard pattern. The vowel part performs the matching by expanding and contracting the frame, so that the characteristic of the temporal change of the local spectrum of the consonant part and the characteristic of the global spectrum of the vowel part can be appropriately captured without being affected by the utterance speed. Therefore, a speech recognition method with high recognition performance can be realized. The amount of calculation and the number of estimated parameters of the standard pattern are increased by using a statistical distance scale expressed by a linear discriminant function that considers inter-frame correlation as a single vector for matching the input voice with the standard pattern. Without this, the recognition rate can be improved. Also, the amount of calculation is doubled, but the frames are treated independently, and instead, a dynamic distance parameter, which is the time variation of the characteristic parameter, is used together and a statistical distance scale expressed by a linear discriminant function is used. , The recognition rate can be improved.

Second, the present invention further improves speech recognition performance by treating a time pattern as an independent vector for each frame and using a statistical distance scale represented by a quadratic discriminant function. it can. When a dynamic feature parameter, which is the amount of change of the feature parameter over time, is used together, the speech recognition performance can be further improved.

Thirdly, the present invention further recognizes syllables and voice segments such as CV (consonant + vowel), VC (vowel + consonant), VCV (vowel + consonant + vowel) or CVC (consonant + vowel + consonant). By combining them, it is possible to realize a highly accurate speech recognition method in which the vocabulary to be recognized can be easily changed.

Also, by introducing the word spotting function, a highly practical recognition device that is robust against noise can be realized.

As described above, the present invention is a practically effective method, and its effect is great.

[Brief description of the drawings]

FIG. 1 is a flowchart showing the flow of processing according to a first embodiment of the present invention;

FIG. 2 is a conceptual diagram illustrating a method of creating a standard pattern in the first embodiment.

FIG. 3 is a diagram showing a DP path in the first embodiment.

FIG. 4 is a flowchart showing the flow of processing according to the second embodiment;

FIG. 5 is a flowchart showing the flow of processing according to the third embodiment;

FIG. 6 is a conceptual diagram illustrating a method of creating a standard pattern in the third embodiment.

FIG. 7 is a flowchart showing a processing flow of a conventional example.

FIG. 8 is a conceptual diagram illustrating a method of creating a standard pattern in a conventional example.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 acoustic analysis unit 2 feature parameter extraction unit 3 voice section detection unit 4 standard pattern storage unit 5 DP matching unit 6 distance comparison unit 7 dynamic feature parameter extraction unit 8 kana notation word dictionary 9 syllable standard pattern storage unit 10 time axis linear normal Conversion unit 11 Distance calculation unit

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Katsuyuki Niyada 3-10-1, Higashi-Mita, Tama-ku, Kawasaki-shi, Kanagawa Prefecture Matsushita Giken Co., Ltd. (56) References JP-A-61-84695 (JP, A) JP-A-61-188599 (JP, A) JP-A-59-17600 (JP, A) JP-A-1-298400 (JP, A) JP-A-58-145999 (JP, A) JP-A-1-136197 (JP, A) JP-A-3-249700 (JP, A) JP-A-58-57196 (JP, A) JP-A-5-73087 (JP, A) JP-A-2-23297 (JP, A) JP 5-4678 (JP, B2) JP 5-4679 (JP, B2) JP 1-15076 (JP, B2) Patent 2712586 (JP, B2) Lecture at the Spring Meeting of the Acoustical Society of Japan in 1995 Papers I, 1-Q-7, Katsutoshi Ohtsuki, "Statistical Phonology Phonological feature analysis of "in p. 109-110 (issued March 14, 1995) R. Rabiner, BH. Jiang "Fundamentals of Speech Recognition", 1993, Prentice-Hall, p. 435-439 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-17/00

Claims (7)

(57) [Claims] 1. A P-number for each frame with respect to the input speech (P is a positive integer) the input obtained by extracting feature parameters
Speech and N types (N is a positive integer) of word sounds to be recognized
From the beginning to the end of the training word voice data belonging to each voice
Until the consonant section is continuously centered on the reference frame.
For vowel parts, linearly expand and contract the vowel section of each data
Makes the whole word J-frame (J is a positive integer) nonlinear
And P (P is a positive integer) for each frame
P obtained by extracting symbol parameters and arranging them in chronological order
The consonant part is collated continuously by using a statistical distance scale, and the vowel part is collated in a time-matched manner by expanding and contracting , using a J-dimensional vector and a word voice standard pattern created in advance. A speech recognition method comprising: obtaining a similarity between a voice and each word voice standard pattern; and using a word voice name corresponding to the word voice standard pattern having the highest similarity as a recognition result. 2. The method according to claim 1, wherein P (P is a positive integer) characteristic parameters are extracted for each frame of the input speech.
Speech and N types (N is a positive integer) of word sounds to be recognized
From the beginning to the end of the training word voice data belonging to each voice
Until the consonant section is continuously centered on the reference frame.
For vowel parts, linearly expand and contract the vowel section of each data
Makes the whole word J-frame (J is a positive integer) nonlinear
And P (P is a positive integer) for each frame
J obtained by extracting symbol parameters and arranging them in temporal order
Using a statistical distance scale, the consonant part is continuously matched with the word speech standard pattern created in advance using the P-dimensional vectors, and the vowel part is expanded and contracted to match in time, A speech recognition method, wherein a similarity between an input speech and each word speech standard pattern is obtained, and a word speech name corresponding to the word speech standard pattern having the highest similarity is used as a recognition result. 3. A statistical distance measure is, speech recognition method of claim 1 or 2, wherein the represented by a linear discriminant function, such as the distance based on Bayesian decision in common the covariance matrix. 4. The speech recognition method according to claim 2 , wherein the statistical distance scale is represented by a quadratic discriminant function such as a distance based on Bayesian judgment or a Mahalanobis distance. 5. A method according to claim 1 to the speech recognition method according to any one of 4, characterized in that the recognition target single syllable of Japanese. 6. vowel unit dynamic programming speech recognition method according to any one of claims 1 to 5, characterized in that matching aligned (DP method) than the time. 7. Performing continuous DP matching using a statistical distance scale based on posterior probabilities to detect a speech section of an unknown input speech and extract a speech portion from a sufficiently long section including noise. claims 1 to methods speech recognition according to any one of 6, characterized by having a recognized word spotting function with.