JP2879989B2 - 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 word voice uttered by an unspecified speaker.

[0002]

2. Description of the Related Art Conventionally, as a method for recognizing a word voice uttered by an unspecified speaker, as shown in FIG. 7, an input voice of a word is analyzed by an acoustic analysis unit 71 to extract feature parameters. A method is known in which this is matched with a word standard pattern 72 created by a number of speakers in advance, and a word similarity is calculated and recognized by a word recognition unit 73. For example, "Unspecified speakers using the word spotting method
Speech Recognition Device for Minority Words "(IEICE SP8
8-18).

In this method, voice data actually produced by 330 speakers is used to create a word standard pattern for recognition of an unspecified speaker. Humans visually cut out voice sections of voice data in which 330 people uttered 10 numbers with reference to spectral waveforms and the like, and obtained a time series of characteristic parameters (LPC cepstrum coefficients) obtained for each analysis time. The voice data is compressed linearly so that the utterance time determined for each word is obtained.
The word standard pattern is created based on the absolute values of the data for 0 persons. By performing matching between the unknown input speech and the standard pattern created in this way using the Mahalanobis distance, which is a statistical distance scale, speech recognition of an unspecified speaker is enabled. This method is based on the idea of statistically absorbing spectral variations of unspecified speakers by comparing and comparing with word standard patterns using a statistical distance scale. The creation of the standard pattern requires data produced by several hundred or more speakers for one recognition word.

As another speech recognition method using a word standard pattern, there is a method using a multi-standard pattern.
This method analyzes a large amount of data, selects multiple representative ones from them, and deals with the spectrum variation of unspecified speakers by comparing multiple word standard patterns with unknown inputs. It is assumed that. Even in this method, it is necessary to collect and analyze data of several hundreds in order to create a plurality of word standard patterns.

As another method for recognizing the word voice of an unspecified speaker, there is a method based on phoneme recognition as shown in FIG. The acoustic analysis unit 81 analyzes the input voice of the word and extracts feature parameters for each analysis time (frame).
The segmentation unit 82 divides the input voice into a vowel section and a consonant section. The phoneme standard pattern 83 is created in advance for each phoneme by acoustically analyzing voice data uttered by many speakers. Next, in the phoneme recognition unit 84,
The vowel section matches the vowel phoneme standard pattern and the consonant section matches the consonant phoneme standard pattern, and recognizes the phoneme to obtain a phoneme symbol string for the input speech. The obtained phoneme symbol string is collated with a word dictionary 86 in which the phonemes are written by the word recognition unit 85, and the similarity of the words is calculated and recognized. This method has an advantage that the word dictionary can be registered as a phoneme notation as compared with the above-described method, so that it is not necessary to collect and analyze a large amount of data for creating a word standard pattern, and the word dictionary can be easily changed. However, since a phoneme is used as a basic unit of recognition, information of a part that changes with time from phoneme to phoneme is not used, and there is a limit in recognition rate.

[0006]

As described above, in the former speech recognition method using the word standard pattern, an enormous amount of work such as data collection and speech segmentation is required to create the word standard pattern of the speech to be recognized. Therefore, there is a problem that the speech to be recognized cannot be easily changed.

[0007] The latter method of using phonemes as a basic unit of recognition has a problem in that the element of temporal change from phonemes to phonemes is lost and there is a limit in increasing the recognition rate.

The present invention solves such a conventional problem, and makes it possible to recognize an unspecified speaker's voice by using recognition target voices uttered by one to several minority speakers.
It is an object of the present invention to provide a voice recognition method that can easily change a voice to be recognized and can obtain a high recognition rate.

[0009]

According to the present invention, in order to attain the above object, at the time of creating a dictionary, one or a few minor speakers utter the recognition target speech, and the resulting speech is subjected to acoustic analysis to obtain characteristic parameters. This time, this is matched with a phoneme standard pattern created by a large number of speakers in advance to obtain a similarity vector for each frame, and the obtained similarity vector is set so that a vector having a large similarity 0 is larger. performs similarity enhancement and normalization, may be registered series in the dictionary when the obtained similarity vector, at the time of speech recognition, the similarity vector in the same manner as in the dictionary registration from an input speech unspecified speaker
In this method, a time series of the tor is obtained and matching with a dictionary is performed, and a recognition target voice having the highest similarity is output as a recognition result.

[0010]

Therefore, according to the present invention, an input speech is analyzed to extract feature parameters, and a similarity vector is determined for each frame from a similarity calculation with a phoneme standard pattern created from speech data of many speakers. By applying an enhancement function to this similarity vector and normalizing it for each frame, 1
By simply registering, as a dictionary, the similarity vector time series of the voices uttered by one or several minority speakers, the voice of the unspecified speaker can be accurately recognized.

[0011]

Hereinafter, embodiments of the present invention will be described.
Before that, the basic concept of the present invention will be described.

The voice of a person is emitted as vibration of the vocal cords in a voiced sound, and the vibration sound is generated in the larynx, pharynx, and so-called articulatory organs.
While passing through the vocal tract formed by the tongue, chin, lips, etc., it undergoes various modulations and is output as speech from the mouth. Therefore,
The phonological properties such as a, i, c, ... are given as the shape of the vocal tract. For unvoiced sounds, the sound source may not be the vocal cords, but phonologicality is also determined by the shape of the vocal tract. However, the shapes and dimensions of the throat, tongue, teeth, chin, lips, and the like that form the vocal tract vary slightly from person to person, and the size of the vocal cords also varies with gender and age. For this reason, the voice of each person differs. That is, the difference in voice due to the difference between persons is largely due to the difference in articulatory organs.

On the other hand, when a voice is uttered as a word or a sentence, not as a phoneme such as a, i, c,..., The shape of the vocal tract changes with time. That is, words are formed by temporal changes in the vocal tract. For example, if you say "red" (akai), the vocal tract changes from a vocal utterance / a / with a narrow back of the tongue to a plosive / k / with a closed larynx and sudden opening. Then, after returning to the / a / shape again, gradually move the tongue to the labial side and move to / i / with the mouth closed. Such a change pattern of the vocal tract is determined by the words to be uttered, and it is considered that there is little difference depending on the person.

As described above, when speech as words is divided into a static vocal tract shape and its temporal change, it is assumed that only the former differs from speaker to speaker, and the latter is considered to have a small difference between speakers. Can be. Therefore, if the difference based on the static vocal tract shape difference can be normalized in some way,
An unspecified speaker can be recognized.

Incidentally, the difference in the shape of the vocal tract is expressed as a difference in the frequency spectrum in the emitted voice signal. The simplest method of normalizing the frequency spectrum between speakers is to match phonemes, syllables, and the like with a short-time voice standard pattern and classify them. If a general-purpose standard pattern created for an unspecified speaker is used, it is possible to obtain similarity information that is not largely affected by differences between speakers. That is, converting a spectrum into similarity information by pattern matching is equivalent to reducing differences between speakers.

On the other hand, since the change pattern of the vocal tract has little difference among speakers, it is sufficient to use information of one to several minor speakers. Therefore, if utterances of a few speakers such as words and phrases are registered in the dictionary as time patterns of similarity information, it is a dictionary for unspecified speakers.

According to the present invention, based on such a concept, 1
A feature parameter obtained by analyzing recognition target voices uttered by several speakers from a name is matched with n kinds of standard patterns created by a large number of speakers in advance for each frame as an analysis time. A time series of an n-dimensional similarity vector is obtained, and this similarity vector is passed through an enhancement function that emphasizes higher similarities, normalized for each frame, and the n-dimensional similarity vector The time series is registered as a dictionary.
When recognizing the input voice, the input voice is also matched with n kinds of standard patterns, and is normalized for each frame through the same emphasis function as that at the time of dictionary registration. The speech recognition of an unspecified speaker is performed by comparing a time series of vectors with the dictionary.

According to the present invention, n types of phonemes, syllables, etc., created by a large number of speakers, are first obtained by analyzing feature parameters obtained by analyzing voices uttered by a small number of one to several speakers. Is obtained for each unit time (for each frame). Since this similarity is a result of matching with a general-purpose standard pattern created by a large number of speakers, the relative relationship between n kinds of similarity values is hardly affected by personality. Therefore, it is effective for an unspecified speaker if the relative relationship of the similarity for each unit time is used as a parameter.

In order to improve the recognition rate, it is only necessary to emphasize a portion contributing to recognition. Therefore, this similarity vector is set such that a portion having a high similarity is larger and a portion having a low similarity does not contribute to the recognition. Through the emphasis function to make such a small value.

Further, in order to equally capture the feature of the relative relationship of the similarity in one frame over the entire voice section, the similarity vector is normalized for each frame. The time series of the n-dimensional similarity vector obtained in this way is registered as a dictionary.

Next, when recognizing the input speech, a time series of n-dimensional similarity vectors prepared as a dictionary and a time series of similarity vectors obtained from the input speech by the same procedure as that for creating a dictionary are used. Collate. Thereby, the voice of the unspecified speaker can be recognized by the dictionary created by a small number of speakers. In the present invention, a higher recognition rate can be obtained because a plurality of candidates are used instead of using only the most reliable one as the similarity per unit time.

Since any word can be described by a combination of phonemes and syllables, a standard pattern of n kinds of phonemes and syllables is created once, so that the same pattern is always obtained even if the recognition target speech is changed. Can be used. To change the recognition vocabulary to change the dictionary so that other voices can be recognized, only a few speakers need to speak. Therefore, it is possible to perform voice recognition of an unspecified speaker by a simple procedure, and to realize a recognition device that is flexible with respect to vocabulary change and the like.

FIG. 1 shows a first embodiment of the present invention.
This will be described with reference to FIG. In FIG. 1, 1 is an acoustic analysis unit, 2 is a feature parameter extraction unit, 3 is a standard pattern storage unit, 4 is a similarity calculation unit, 5 is a similarity normalization unit by emphasis, 6 is a parameter time series creation unit, 7 is a dictionary storage, 8
Is a pattern matching unit.

Next, the operation of the present embodiment will be described for the case where the voice of one speaker is registered in a dictionary. That is, in this embodiment, first, a dictionary is created using the recognition target voice uttered by one person as an input voice, and at the time of recognition, the input voice of the unspecified speaker is recognized using the dictionary.

In FIG. 1, when an input voice is input, a frame (analysis time is 1
A linear prediction coefficient (LPC) is determined every frame (10 msec).

Next, the feature parameter extraction unit 2 performs LPC
Cepstrum coefficients (9 from C0 to C8) are obtained.

The standard pattern storage unit 3 stores n types of phoneme standard patterns created in advance from data uttered by many speakers. In this embodiment, n = 20 and / a /, / o /, / u /, / i /, / e /, / j /, / w /, / m /, / n /, /

[0028]

[Outside 1]

/, / B /, / d /, / r /, / z /, / h /, / s /, / c /, / p /, / t /,
Use 20 phoneme standard patterns of / k /. The phoneme standard pattern is created by accurately detecting the characteristic portion of each phoneme (a temporal position at which the characteristics of the phoneme is well expressed) by visual observation, and using a time pattern of feature parameters centered on the feature frame. In this embodiment, the time pattern is eight frames before the feature frame and three frames after the feature frame, for a total of one.
LPC cepstrum coefficient for two frames (C0 to C8)
Form a parameter time series. Then, a parameter time series is extracted from a large amount of data uttered by many people, and an average vector μp of each element and a covariance matrix 間 の between the elements are obtained as a standard pattern. As described above, the phoneme standard pattern used in the present embodiment uses feature parameters of a plurality of frames, and is characterized in that the standard pattern is created in consideration of the temporal movement of the parameters.

The similarity between the 20 types of phoneme standard patterns and the feature parameters (LPC cepstrum coefficients) obtained by the feature parameter extraction unit 2 is calculated by the similarity calculation unit 4 for each frame. That is, matching is performed with the standard pattern while shifting the input one frame at a time.
A time series of similarity such as In this embodiment, a Mahalanobis distance in which a covariance matrix is shared is used as a distance measure for similarity calculation. The Mahalanobis distance d _p for calculating the similarity between the input and the standard pattern of the phoneme p is represented by the following (Equation 1). Here, x is a vector composed of feature parameters for 12 frames, which is an input time pattern.

[0031]

(Equation 1)

Here, assuming that the covariance matrix 共通_p is common to each phoneme, it can be expanded into the following simple equation (Equation 2). Let Σ be the common covariance matrix.

[0033]

(Equation 2)

In this embodiment, the above (Equation 2), which requires a small amount of calculation, is used. a _p and b _p are standard patterns for the phoneme p, and are stored in the standard pattern storage unit 3 in advance. A vector having similarities to the 20 types of phoneme standard patterns obtained in this way (shaded portions in FIG. 2) is referred to as a similarity vector.

The similarity emphasizing / normalizing unit 5 passes the similarity vector obtained by the similarity calculating unit 4 through an emphasizing function for emphasizing higher similarities, and the maximum value is 1 for each frame. Normalize so that the minimum value is 0. This is performed over the entire frame, and the parameter sequence creation unit 6 creates a time series of similarity vectors.

The similarity emphasis / normalization unit 5 converts the similarity vector obtained for each frame as follows. First, the 20 elements of the similarity vector are arranged in descending order (it is assumed that the larger the value is, the more similar the phoneme standard pattern is), and the first similarity is 1 and the k-th similarity is 0. The values of similarity from the first place to the k-th place are linearly converted to 1 to 0 so that From the (k + 1) th place to the twentieth place, all are set to 0, and a new similarity vector including 20 elements is obtained. That is, the similarity vector is defined as a = (a
_{1, a 2, ..., a} i, ..., When a _20), the enhancement function F is expressed by the following equation (3). Here, M is the maximum value of the similarity, and M _k is the value of the k-th similarity.

[0037]

(Equation 3)

By passing such a function F, the degree of similarity to a higher-order phoneme is emphasized. Also, since the values of M and M _k differ from frame to frame, F
Although (a _i ) also differs for each frame, the maximum value in the frame is always 1 and the minimum value is 0, and is normalized for each frame. Since normalization is performed for each frame in this manner, since the phoneme standard pattern is created from the time patterns of the characteristic parameters around the feature frame, the degree of similarity to any phoneme standard pattern in the transition (transition) of a phoneme is determined. The reason for this is that if the whole becomes small and normalization is not performed for each frame, the feature of the relative relationship of the similarity in the transition portion of the phoneme will be underestimated. Therefore, normalization is performed for each frame so that the relative relationship of the similarity can be equally handled over the entire voice section.

In this way, a new similarity vector is obtained over the entire voice section, and the parameter time series creating section 6 creates a time series of similarity vectors. Figure 3 is an example of a time series of similarity vector after emphasizing-normalized time series of similarity vector of Figure 2. Here, paying attention to the similarity vector in the hatched portion, the similarity of the phoneme / a / having the maximum similarity in FIG. 2 is 1, and the k-th place (for example, phoneme / p /).
Is linearly converted from the first place to the k-th place so that the similarity of the first place becomes zero, and all small values below the (k + 1) -th place are set to zero.

The procedure up to this point is the same for both dictionary creation and recognition.

At the time of creating a dictionary, one uttered speech to be recognized is input as input speech, and the obtained time series of similarity vectors is registered in the dictionary storage unit 7.

At the time of recognition, the time series of the similarity vector is obtained from the input speech in the same manner as when the dictionary is created, and the pattern matching unit 8 matches the time series of the similarity vector in the dictionary storage unit 7 to obtain the highest score. A large dictionary item is used as a recognition result. In this embodiment, DP matching is performed as a matching method. (Equation 4) shows an example of a recurrence formula for performing DP matching. Here, the length of the dictionary is J frame, the input length is I frame, the distance function between the i-th frame and the j-th frame is l (i, j), and the cumulative similarity is g
(I, j).

[0043]

(Equation 4)

In this embodiment, the Euclidean distance is used as a distance measure of the distance function l (i, j). Let a = (a ₁ , a ₂ ,...,
a ₂₀ ), the similarity vector in the j frame of the dictionary is represented by b
= (B ₁ , b ₂ ,..., B ₂₀ ), l (i, j) when the Euclidean distance is used is represented by (Equation 5).

[0045]

(Equation 5)

As described above, the similarity to the upper phoneme is emphasized, and the Euclidean distance between the vectors in which the similarity to the lower phoneme is uniformly set to 0 is obtained, thereby emphasizing the movement of the similarity to the upper phoneme. Therefore, the movement of the similarity to the lower phoneme can be ignored. Further, by performing normalization for each frame, it becomes possible to treat the feature of the relative relationship of the similarity in the transition part of the phoneme with the same weight as that around the feature frame. Therefore, by performing such emphasis and normalization, a high recognition rate can be obtained.

Next, a case where the voices of two or more speakers are registered in the dictionary will be described. The recognition method is the same as in the case where a dictionary is registered from one utterance described above.
First, a method of adjusting the time of the same voice uttered by a plurality of speakers by DP matching and registering the same as a dictionary will be described. Next, the same voice uttered by a plurality of speakers will be registered in the dictionary as a multi-standard pattern. The method will be described.

When there are two speakers, DP matching is performed and time matching is performed in the same manner as in the case of recognizing the same two voices. The time alignment will be described with reference to FIG. FIG. 4 is an example in which “red” (akai) and two speakers uttered. Since the utterance time length differs depending on the speaker, DP matching is performed between the same recognition target voices of the two speakers, and the DP path is inversely traced from the result to perform time matching. By performing time matching, sections of the same phoneme (/ a /, / k /, / a /, / i /) are matched. Then, an average value of each similarity is obtained between the frames that are temporally matched, and the time series is registered as a dictionary.

That is, the first speaker 1 shown by hatching in FIG.
The i-th frame and the j-th frame of speaker 2 are temporally aligned
In this case, the similarity vector of the i-th frame of the speaker 1 is represented by c =
(C ₁, C_Two, ..., c₂₀), And the j-th frame of speaker 2 is e =
(E₁, E_Two, ..., e₂₀), F = ((c₁
+ E₁)) / 2, (c_Two+ E_Two) / 2, ..., (c₂₀+
e₂₀) / 2), and finds the similarity vector f in the dictionary i
Register as a frame similarity vector. Like this
And improve the accuracy of the dictionary and increase the recognition rate
Obtainable.

Next, when voices uttered by a plurality of speakers are registered in the dictionary as a multi-standard pattern, a plurality of similarity vector time series of voices uttered by a plurality of speakers are registered as the dictionary as recognition target voices. . In this case, it is assumed that the dictionary item is recognized even if it is recognized in any dictionary among the standard patterns registered for each dictionary item.

However, when a dictionary is created by the utterances of two or more speakers, in order to reduce the gender difference in the dictionary pattern, the dictionary is created by one male and one female or almost the same number of male and female utterances.

Next, a speech recognition experiment using this embodiment and the results thereof will be described. The experiment used data of 20 people who uttered 212 words, and one of the 20
The data which uttered two words is registered as a dictionary, and the other 19 words are registered.
The recognition was performed by a method of recognizing 212 words uttered. As a result of the experiment, a recognition rate of 88.5% was obtained. On the other hand, the recognition rate when the time series of the similarity obtained by matching with the phoneme standard pattern is used as is and normalization is not performed for each frame is 82.1.
%, Indicating that the emphasis effect of the similarity greatly contributes to the recognition rate.

When a dictionary is created by uttering one speaker,
The recognition rate of that speaker and that of the opposite sex is 86.0% on average, which is 5 times higher than the average recognition rate of 91.4% for the voice of the same sex speaker.
% Lower. Therefore, if a time series pattern obtained by averaging the time series of the similarity vectors obtained from the recognition target voices uttered by two persons, one for each man and woman, is used as a dictionary, the gender difference is eliminated, so that 93.4% High recognition rate was obtained. In a method using a multi-standard pattern in which two voices uttered by one male and one female each were registered as a dictionary without averaging, a recognition rate of 93.2% was obtained.

As described above, the input speech is analyzed to extract the characteristic parameters, and the similarity vector is obtained for each frame from the similarity calculation with the phoneme standard pattern created by the speech data of many speakers. By applying an enhancement function to the degree vector and normalizing for each frame, the similarity vector of the input voice can be registered simply by registering a time series of the similarity vector of the voice uttered by one or several minor speakers as a dictionary. The voice of the unspecified speaker can be accurately recognized by the DP matching between the time series and the dictionary.

Next, the similarity vector is passed through an emphasis function such as an exponential function that emphasizes a portion having a large similarity value, and a regression coefficient representing a time change amount thereof is used together to obtain a correlation cosine. A second embodiment of the present invention, in which recognition is performed by the following, will be described with reference to FIG.

In FIG. 5, reference numeral 11 denotes an acoustic analyzer, 12 denotes a feature parameter extractor, 13 denotes a standard pattern storage, 1
4 is a similarity calculation unit, 15 is a similarity enhancement unit, 16 is a similarity normalization unit, 17 is a regression coefficient calculation unit, 18 is a regression coefficient normalization unit, 19 is a parameter time series creation unit, 20 is The dictionary storage unit 21 is a pattern matching unit.

Also in the second embodiment, similarly to the first embodiment, the input voice is analyzed by the acoustic analysis unit 11 to obtain the characteristic parameters by the characteristic parameter extracting unit 12 and stored in the standard pattern storage unit 13 in advance. The registered phoneme standard pattern is matched for each frame, and a time series of similarity vectors is obtained by the similarity calculation unit 14.

Next, in the similarity emphasizing unit 15, the similarity calculated by the similarity calculating unit 14 is passed through an emphasizing function G, which is an exponential function, so that the similarity having a larger value is converted to be larger. This emphasis function G calculates the similarity vector of the input speech as a = (a ₁ , a ₂ , a ₃ ,..., A _i ,.
₂₀ ), it is expressed by (Equation 6).

[0059]

(Equation 6)

Α and β are constants common to all phonemes and all frames, and a similarity vector is newly calculated for all frames by using this equation.

Further, the similarity normalizing section 16 normalizes the n-dimensional similarity vector to a size of 1 for each frame to create a new similarity vector. When this is expressed by an equation, it becomes as shown in (Equation 7).

[0062]

(Equation 7)

Here, the similarity vector emphasized by the emphasis function G is defined as a ′ = (a ₁ ′, a ₂ ′, a ₃ ′,.
_{a i ', ..., a 20} ' and), was the magnitude of 1 vector _{a "= (a 1",} a 2 ", a 3", ..., a i ", ...,
a ₂₀ ″). By setting the magnitude of the similarity vector for each frame to 1, the feature of the relative function of the similarity over the entire voice section can be treated equally.

Next, the regression coefficient calculation unit 17 obtains a regression coefficient (n) which is a temporal change amount of the similarity with respect to each normalized time series of the similarity for each frame. The regression coefficient is
It is the inclination of each similarity to each phoneme in the time direction. That is, for example, first, in the time series of the similarity of the phoneme / a / with respect to the standard pattern, the slope (similarity of similarity) of the least square approximation line of the similarity values of two frames before and after a certain frame (similarity values of five frames in total). Temporal change). This is shown in (Equation 8).

[0065]

(Equation 8)

Here, xt = (t = 1, 2, 3,...) Represents a time series of similarity to phoneme / a /, and K (/ a /) is a regression coefficient of phoneme / a / at time t + 2. It is. This is obtained for 20 similarities for each phoneme, and further for every frame for every frame to obtain a time series of regression coefficient vectors.

Next, the regression coefficient normalization unit 18 normalizes the regression coefficient vector to a size of 1 for each frame, similarly to the similarity.

Then, the parameter time series creating unit 19 converts both the time series of the magnitude 1 n-dimensional similarity vector emphasized by the exponential function and the time series of the magnitude 1 n-dimensional regression coefficient vector. Parameter time series.

The procedure up to this point is the same for both dictionary creation and recognition. At the time of dictionary creation, as in the first embodiment, first, a dictionary is created using one uttered recognition target speech as input speech. During recognition, the input speech of the unspecified speaker is recognized using the dictionary.

At the time of creating a dictionary, the parameter time series obtained by the parameter time series creating unit 19 is registered in the dictionary storage unit 20. A dictionary may be created and registered from the same voice uttered by two or more minority speakers in the same manner as described in the first embodiment.

At the time of recognition, the pattern matching unit 21 performs DP matching between the parameter time series obtained by the same method as that at the time of dictionary registration and the time series of parameters in the dictionary storage unit 20 using the correlation cosine. The dictionary item having the highest similarity is regarded as the recognition result. The recurrence formula for performing DP matching is the same as (Equation 4) used in the first embodiment, but the present embodiment uses the correlation cosine as a distance measure of the distance function 1 (i, j). The distance function 1 (i, j) using the correlation cosine in combination with the regression coefficient is represented by (Equation 9).

[0072]

(Equation 9)

[0073] However, the similarity vector of the i-frames of the input speech _{a = (a 1, a 2} , ..., a 20), the regression coefficient vector _{c = (c 1, c 2} , ..., c 20), Dictionary B = (b ₁ , b ₂ ,..., B)
₂₀ ), and the regression coefficient vector is d = (d ₁ , d ₂ ,..., D ₂₀ )
And w is a mixture ratio of the similarity and the regression coefficient.
4 to 0.6 is good. In fact, already the similarity vector,
Since both the regression coefficient vector and the regression coefficient vector are normalized to the size 1, it is only necessary to calculate the inner products and add up the weights with w: (1-w). That is, (Expression 10) is obtained.

[0074]

(Equation 10)

Next, a speech recognition experiment using the second embodiment and the results thereof will be described. The experiment used data of 20 people who uttered 212 words, and one of the 20
The data which uttered 12 words is registered as a dictionary, and the other 1
The method was performed by recognizing 212 words uttered by nine people.
When the regression coefficient vector in the second embodiment is used together with the first embodiment shown in FIG. 1, a word recognition rate of 90.3% was obtained. This is 88.5 without regression coefficients.
% Is improved by 1.8%. In addition, the second shown in FIG.
When the regression coefficient was used in combination with the method of the embodiment, the result was 91.6%, and the recognition rate was further improved by 1.3%.

Further, data obtained by averaging the word voices uttered by a total of two speakers, one for each man and woman, is registered as a dictionary, and when the remaining 18 word voices are evaluated, a high recognition rate of 95.9% is obtained. Obtained.

As described above, the emphasis function is applied to the similarity,
By performing pattern matching using the correlation cosine in combination with the regression coefficient, the recognition rate is improved.

In the second embodiment, the regression coefficient is
The similarity vector was obtained by normalizing the similarity vector to a size of 1 by emphasizing it with an exponential function, but as shown in FIG. May be obtained for the similarity series emphasized by the exponential function in the emphasizing unit 15.

[0079]

As described above, according to the present invention, the similarity is calculated by calculating the similarity between the feature parameters obtained by analyzing the voice and n kinds of standard patterns created by many speakers in advance. A sequence is obtained, and the similarity is emphasized so that the larger the similarity becomes larger for each frame through an enhancement function, and the n-dimensional similarity vector or the n-dimensional similarity vector and its n-dimensional regression coefficient vector are subjected to speech recognition. By registering the recognition target speech uttered by one or a few speakers as a dictionary by using the feature parameters for Can be.

Therefore, it is extremely easy to create a dictionary, and when it is desired to change the recognition target voice, the voice data uttered by one or a small number of speakers can be updated simply by registering it as a dictionary.

Further, by normalizing each frame through the enhancement function, a higher recognition rate can be obtained.

Further, by creating a dictionary from voice data uttered by the same number of male and female minority speakers, a higher recognition rate can be obtained.

As described above, the present invention greatly contributes to the improvement of the performance of the speech recognition apparatus for unspecified speakers and the improvement of the flexibility for application to various uses.

[Brief description of the drawings]

FIG. 1 is a functional block diagram for implementing a voice recognition method according to a first embodiment of the present invention;

FIG. 2 is a time series diagram showing an example of a time series of similarity vectors in the embodiment.

FIG. 3 is a time series diagram showing an example of a time series of a similarity vector after emphasis and normalization in the embodiment.

FIG. 4 is a schematic diagram showing an example of time alignment for registered voices of two speakers in the embodiment.

FIG. 5 is a functional block diagram according to a second embodiment of the present invention.

FIG. 6 is a functional block diagram showing a modification of the second embodiment of the present invention.

FIG. 7 is a functional block diagram showing an example of a conventional word speech recognition method.

FIG. 8 is a functional block diagram showing another example of the conventional word speech recognition method.

[Explanation of symbols]

REFERENCE SIGNS LIST 1 acoustic analysis unit 2 feature parameter extraction unit 3 standard pattern storage unit 4 similarity calculation unit 5 similarity enhancement / normalization unit 6 parameter time series creation unit 7 dictionary storage unit 8 pattern matching unit 11 acoustic analysis unit 12 feature parameter extraction Unit 13 Standard pattern storage unit 14 Similarity calculation unit 15 Similarity enhancement 16 Similarity normalization unit 17 Regression coefficient calculation unit 18 Regression coefficient normalization unit 19 Parameter time series creation unit 20 Dictionary storage unit 21 Pattern matching unit

Continuation of the front page (56) References JP-A-1-216397 (JP, A) JP-A-60-200296 (JP, A) JP-A-2-248999 (JP, A) JP-A-62-66300 (JP, A) JP-A-62-145297 (JP, A) JP-A-60-202488 (JP, A) JP-A-60-209794 (JP, A) JP-B-62-17760 (JP, B2) JP-B-sho 61-54240 (JP, B2) IEEE Transactions on Acoustics, Speech and Signaling, Vol. ASSP-34, no. 1, p. 52-59 (1986) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 531 G10L 3/00 521 JICST file (JOIS)

Claims (8)

(57) [Claims] [Claim 1] At the time of dictionary creation, the recognition target voice uttered a small number of speakers of several people from one person, asked Me the m feature parameters for each is the analysis time frame, sub et al beforehand a number of story The matching is performed with n kinds of standard patterns created by the user to obtain n similarities for each frame, and the obtained n-dimensional similarity vector is determined such that the larger similarity vector becomes larger. Normalized for each frame through the enhancement function
When registering a time-series pattern of the n-dimensional similarity vector created as a dictionary as a dictionary and recognizing the input voice,
M features obtained for each frame by acoustic analysis of input speech
A parameter is matched with the n kinds of standard patterns to obtain an n-dimensional similarity vector, and a time series of the similarity vector is created through the same emphasis function as that at the time of creating a dictionary. Is compared with the time series pattern of the similarity vector registered in the dictionary.
Speech recognition method and, and recognizes the input speech. 2. An n-dimensional similarity vector with respect to each dimension of a time series of n-dimensional similarity vectors to which an enhancement function has been applied is determined for each frame by n pieces. 2. The speech recognition method according to claim 1, wherein a time-series pattern is created by using both the dimensional vector and the n-dimensional vector of similarity. 3. When comparing an input voice with a dictionary, a similarity vector or a similarity vector of the input voice and a time variation vector thereof normalized for each frame, and a similarity vector or a similarity vector of the dictionary voice 3. The voice recognition method according to claim 1, wherein the matching is performed based on a distance between the time and a time variation vector. 4. As the enhancement function, the first to k-th elements when elements of the similarity vector are arranged in descending order are selected and used, and a function having the lowest value in the (k + 1) -th and lower elements is used to rank the elements. 3. The speech recognition method according to claim 1, wherein weighting is performed, and a Euclidean distance is used as a distance measure for comparing the input speech with the dictionary. 5. The method according to claim 1, wherein an exponential function is used as an emphasis function such that a larger value is larger, and a correlation distance is used as a distance measure for collation between the input speech and the dictionary. Voice recognition method. 6. An n-dimensional similarity vector or a time series of an n-dimensional similarity vector and an n-dimensional time variation vector obtained by two or more speakers uttering the same recognition target voice and analyzing each of them. And performing time matching between speakers by DP matching, calculating an average value of each similarity between frames matched in time, and registering a time-series pattern of the average value in a dictionary. 3. The speech recognition method according to 1 or 2. 7. The same recognition target voice is uttered by two or more speakers and analyzed to obtain a plurality of n-dimensional similarity vectors or a plurality of time series of an n-dimensional similarity vector and an n-dimensional time variation vector. And registering them as a dictionary and using them as a multi-standard pattern.
Or the speech recognition method according to 2. 8. The speech recognition method according to claim 6, wherein, when creating a dictionary based on the utterances of two or more speakers, the dictionary is created based on one male and one female or approximately the same number of male and female utterances.