JP3291073B2 - 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 speech recognition system for recognizing uttered speech.

[0002]

2. Description of the Related Art A method of recognizing a speech by word spotting, instead of recognizing the entire speech section of a uttered speech without a gap (one character and one phrase), avoids problems such as addition of non-words and pause. It is known that it is suitable for a speech dialogue system and a speech understanding system.

Also, when recognizing a discretely uttered word voice, it is better to use a spotting technique from a pattern including a silent part instead of recognizing the voice section after extracting a voice section based on information such as power. There is an advantage that it is not affected by noise or tongue sound at the time of sound reception.

[0004]

However, spotting has a problem of partial matching. For example, if the words to be recognized include the words "Shin-Yokohama" and "Yokohama", and when the speaker utters "Shin-Yokohama", "Yokohama" is included in the voice, Both "Yokohama" and "Shin-Yokohama" were recognized with a high score (small distance), and there was a problem that it was not possible to identify which of these words. This partial matching is particularly problematic when recognizing numbers. For example,
The words "31" include "30", "3", "1"
There are partially matched words such as "0", "11", and "1".

However, this partial matching has asymmetry. That is, a long word (“Shin-Yokohama” in the above example) may be erroneously recognized as a short word (“Yokohama”), but the converse is not so many. For example, the method in Japanese Patent Application Laid-Open No. 4-230797 utilizes this asymmetry. In other words, a similarity table is created in advance that the similarity of “Yokohama” is high for the input “Shin-Yokohama”, but is not high for the input “Yokohama”. First, a normal collation is performed, and then, the similarity to all the words obtained here is compared with the similarity table described above, and a word having a similar similarity (smallest distance) is regarded as a recognition result. Is what you do.

However, in the prior art, since it is necessary to prepare a similarity table of all words to be recognized in advance, it is effective only for an application in which the vocabulary to be recognized is fixed. For example, there is a drawback that the problem of partial matching cannot be solved with a recognition device that can freely change words to be recognized like a recognition device of a specific speaker system.
Further, there is a disadvantage that the processing is complicated because the collation is performed in two stages.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a matching method which can avoid the problem of partial matching even if the vocabulary to be recognized is changed only by one-stage matching.

[0008]

A technical solution according to the present invention for solving the above problem is to convert an input voice into an input pattern which is a time series of a feature vector, and to provide a first predetermined pattern. The value obtained by subtracting the distance between the feature vector of the input pattern and the feature vector of the standard pattern from the value is defined as similarity, and the sum of the similarity for the series of feature vectors of the standard pattern is calculated as the similarity with the input voice. Deemed,
Recognize input voice. Here, the first value given in advance is determined for each feature vector of the standard pattern, or the feature vector of the standard pattern is divided into a plurality of clusters in advance and determined for each cluster. Further, after matching is performed on an input pattern whose voice content is known, the feature vector of the input pattern and the feature vector of the standard pattern are associated based on the matching path, and the two associated features are compared. When the distance of the vector is larger than the second predetermined value, the first predetermined value is increased for each standard pattern, or the first predetermined value is set for each cluster to which the standard pattern belongs. It is decided to increase the value.

[0009]

According to the present invention, first, the similarity is defined as a value obtained by subtracting the distance between the feature vector of the input pattern and the feature vector of the standard pattern from the first value given in advance. Here, the first value given in advance is determined for each feature vector of the standard pattern, or the feature vector of the standard pattern is divided into a plurality of clusters in advance and determined for each cluster. Then, the input speech is converted into an input pattern that is a time series of feature vectors, and the sum of similarities to the sequence of feature vectors of the standard pattern is regarded as similarity to the input speech, and By recognizing the voice, the problem of partial matching can be avoided by one-stage matching, and words can be correctly recognized. Further, since the information of the vocabulary to be recognized is not required in advance, even a recognition device that changes the vocabulary, such as a specific speaker type speech recognition device, can operate. Further, after matching is performed on an input pattern whose voice content is known, the feature vector of the input pattern and the feature vector of the standard pattern are associated based on the matching path, and the two associated features are compared. When the distance of the vector is larger than the second predetermined value, the first value predetermined for each standard pattern is increased.
Alternatively, the first value predetermined for each cluster to which the standard pattern belongs is set to be large, and the similarity between the feature vector of the input pattern and the feature vector of the standard pattern associated based on the matching path is small. By controlling so as not to take a value, correct speech recognition becomes possible by minimizing the possibility that the standard pattern is not spotted.

[0010]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic block diagram of a speech recognition device according to the present invention. Referring to FIG. 1, the voice recognition device includes a voice input unit 1 such as a microphone or a receiver for inputting voice, a feature extraction unit 2 for converting an input voice signal into a time-series input pattern of a feature vector, A standard pattern storage unit 6 for storing a standard pattern of voice, and a collating unit 3 for collating the input pattern of the extraction result with the standard pattern
And a similarity calculating unit 4 for calculating the similarity between the input pattern and the standard pattern to be compared, and a matching path searching unit 5 for searching for a corresponding position having a large similarity between the input pattern and the standard pattern. are doing.

The feature extraction unit 2 extracts an input pattern using feature quantities such as LPC mel-cepstrum, among various parameters useful for speech recognition. Hereinafter, the input pattern X is represented by X = x ₁ x ₂ ... X _I
Will be represented as follows. (Here, I is the total number of frames of the input pattern.)

The collating unit 3 compares the feature vector of the input pattern extracted by the feature extracting unit 2 with the feature vector of the standard pattern in the standard pattern storage unit 6. Here the standard pattern Y of the word K is represented by a vector column such as _{Y = y 1 y 2 ··· y} J. (Where J
Is the total number of frames of the standard pattern. At this time, the distance between the feature vector x _i of the i-th frame of the input pattern and the feature vector y _j of the j-th frame of the standard pattern Y is described as d (x _i , y _j ). To The definition of distance is city distance,
Various methods such as the Euclidean distance and the Mahalanobis distance are known, and any of them may be used.

[0013] similarity calculator 4 defines a feature vector x _i of the i-th frame of the input pattern, the similarity between the feature vector y _j of the j-th frame of reference pattern Y as follows .

[0014]

(Equation 1)

A method of spotting the standard pattern Y from the input pattern X based on the similarity obtained as described above will be described below. Various methods are known as this method. It is not limited to.

The collation path search section 5 searches in the following procedure. First, an array D (i, j) for storing the accumulated score when the matching is completed up to the j-th feature vector of the standard pattern at the time of the input frame i, and an array B (for storing the start time of this matching path) i, j) are prepared. At the initial time, the array D (i, j) and the array B (i, j) are determined as follows.

[0017]

(Equation 2)

Here, in equation (2), when the large value is (a), B (i, 1) = B (i-1,1), and when the large value is (b), B (i, 1) = i are set respectively. At the intermediate point, the arrays D (i, j) and B (i, j) are determined as follows.

[0019]

(Equation 3)

Here, in equation (3), when the large value is (c), B (i, j) = B (i-1, j), and when the large value is (d), B (i, j) = B (i-1, j-1)

D (i, J) is the word K
, And K is B (i,
J) A recognition result that the frame exists from the frame to the i-th frame is obtained. Although the recognition result has i degrees of freedom, when the number of recognition results is reduced to one, it is sufficient to select i that maximizes D (i, J).

Next, a method of setting the value of S will be described. (1) When the value of S is determined in common for all the standard patterns As a setting method of S, a preliminary experiment is performed, and the value of max D (i, J) for the standard pattern having the same content as the utterance content is positive. And the value of max D (i, J) for a standard pattern having a content different from the utterance content may be set to be negative.

(2) When the value of S is determined for each feature vector of the standard pattern As a setting method, if the value of S for a certain feature vector y _m of a certain standard pattern is described as S _m , then S _m it may be set by adding a positive constant d ₀ on the average D_ave _m distance data and y _m for learning that created the y _m. Here, as an example of determining d ₀ , N can be experimentally determined using equation (4), where N is the number of feature vectors of the standard pattern.

[0024]

(Equation 4)

[0025] Also, most of the learning (e.g. 95%) may be set to the value of S _m as the degree of similarity is positive for the vector y _m.

(3) The feature vector of the standard pattern is divided into a plurality of clusters in advance.
When the value of is determined The cluster division method is to divide for each phoneme or a group of phonemes (vowel, unvoiced fricative, nasal, plosive, etc.)
What is necessary is just to divide into. Here, S for a certain cluster M
When the values described as S _M, with respect to elements belonging m to the cluster M, the average value D_ave _m of the distance data and y _m for learning that created the y _m, a positive constant that this averaged d ₀ (the same method as described above may be used to determine this value) may be added and set. Further, the value of S _M may be set so that most (eg, 95%) of the learning data for the feature vector y _m belonging to the cluster M has a positive similarity to y _m .

Further, the input pattern whose utterance content is known is compared with a vector sequence of a standard pattern having the same content, and then backtracking is performed. Here is the feature vector y _m of the feature vector xi and the standard pattern of the input pattern and with corresponding on the basis of the collation path. Here, the similarity r (x
i, if smaller than the threshold TH value of y _m) is set in _{advance, S m = S m + α} (TH-r (xi, y m))
Sets the value of S for y _m large. Likewise
And divide the feature vector of the standard pattern into clusters
The cluster to which the input pattern y _m belongs is M
If _{you, S M = S M + α} (TH-r (xi, y m)) to
Therefore, the value of S is set large.

Here, the learning coefficient α is set to, for example, about 0.1. As an example of determining TH, N can be experimentally determined using equation (5), where N is the number of feature vectors of the standard pattern. At this time, if TH is determined to be small, the error of "no recognition result" increases for a correct utterance, but erroneous recognition decreases. If TH is determined to be larger, the error of "no recognition result" is reduced, but the number of erroneous recognition is increased. After recognizing such a property, it is necessary to appropriately determine it according to an application example in which the speech recognition device is used.

[0029]

(Equation 5)

FIG. 2 is an example showing the similarity of the partial pattern matching when the input voice is "Shin-Yokohama". FIG. 2A is a diagram illustrating an input voice. FIG.
(B) is a diagram showing the distance between the input voice and the standard pattern “Shin-Yokohama”. FIG. 2C is a diagram illustrating a distance between the input voice and the standard pattern “Yokohama”. As can be seen from this figure, since the similarity to the “new” part, which is the part A in FIG. 2, is a positive value, the matching score for the standard pattern “shin-yokohama” is larger by that amount. "Shin-Yokohama" is correctly recognized.

FIG. 3 is an example showing the similarity of the partial pattern matching when the input voice is "Yokohama". FIG. 2A is a diagram illustrating an input voice. FIG.
(B) is a diagram showing the distance between the input voice and the standard pattern “Shin-Yokohama”. FIG. 2C is a diagram illustrating a distance between the input voice and the standard pattern “Yokohama”. As can be seen from this example, since the “new” part of the standard pattern is compared with the non-speech section or another word that is not “new” in the comparison with the standard pattern of “Shin Yokohama”, the similarity of this part is Becomes a negative value (part B in FIG. 3),
Since the matching score for the standard pattern of “Shin-Yokohama” is small, “Yokohama” is still correctly recognized.

[0032]

As described above, according to the present invention, the problem of partial matching can be avoided by only one-stage matching, and words can be correctly recognized. Further, since the information of the vocabulary to be recognized is not required in advance, even a recognition device that changes the vocabulary, such as a specific speaker type speech recognition device, can operate. According to the related art, the similarity between the feature vector of the input pattern and the feature vector of the standard pattern, which are associated based on the matching path, is a small value (eg, a negative number). , There is a high possibility that this standard pattern will not be spotted, but in the present invention, correct speech recognition can be performed by resetting the value of S so as to increase the similarity in such a state. become.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram of a speech recognition device according to the present invention.

FIG. 2 is an embodiment to which the present invention is applied.

FIG. 3 is another embodiment to which the present invention is applied.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Voice input part, 2 ... Feature extraction part, 3 ... Matching part, 4 ... Similarity calculation part, 5 ... Matching path search part, 6 ... Standard pattern storage part.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/10

Claims (6)

(57) [Claims] 1. A speech input unit for inputting speech, a feature extraction unit for converting speech input from the speech input unit into an input pattern which is a time series of feature vectors, and a feature extracted from the feature extraction unit. A matching unit that calculates the distance between the feature vector of the input pattern and the feature vector of the standard pattern, and calculates the difference between the first value given in advance and the distance calculated by the matching unit. Speech recognition, wherein the similarity calculated and accumulated for a series of feature vectors of a standard pattern for a word to be recognized is accumulated and used as the similarity of the word to be recognized. method. 2. The speech recognition method according to claim 1, wherein the first value given in advance is a value common to feature vectors of all standard patterns. 3. The speech recognition method according to claim 1, wherein the first value given in advance is determined for each feature vector of a standard pattern. 4. The speech recognition method according to claim 1, wherein the feature vector of the standard pattern is divided into a plurality of clusters in advance, and the predetermined first value is determined for each cluster. . 5. After performing matching on an input pattern whose voice content is known, a feature vector of the input pattern is associated with a feature vector of a standard pattern based on a matching path. 4. The voice according to claim 3, wherein when the distance between the two feature vectors is larger than a predetermined second value, the predetermined first value is increased for each of the feature vectors of the standard pattern. Recognition method. 6. After matching is performed on an input pattern whose utterance content is known, the feature vector of the input pattern and the feature vector of the standard pattern are associated based on the matching path. 5. The speech recognition according to claim 4, wherein when the distance between the two feature vectors is larger than a second predetermined value, the first predetermined value is increased for each cluster to which the standard pattern belongs. method.