JPH071434B2 - Standard pattern creation method - Google Patents

Description

TECHNICAL FIELD The present invention relates to an improvement of a standard pattern creating method in speech recognition.

[Prior Art] Differences in the shape of the vocal organs by speakers and differences in articulation patterns due to vocal habits and habits impede the realization of voice recognition for the voices of unspecified speakers as individual differences in voice. It is a factor. As a method to deal with this, a method of normalizing individual differences of speech, a method of adapting a recognition system to a speaker with a small amount of learning samples, and a method of extracting a feature independent of a speaker from speech and using it for recognition Are being considered.

One of the methods of adapting the recognition system to the speaker by a small amount of learning samples is the paper "Nakajima, Takahashi," Speaker adaptation method in large vocabulary speech recognition ", Proceedings of the Acoustical Society of Japan.
There is a method proposed in "1-1-6, October 1983". This is based on the pseudophonic standard pattern type (SPLIT type in the following, using the frame correspondence between two voices obtained by DP matching between the learning voice uttered by an unknown input speaker and the same category voice uttered by a standard speaker. The standard quantization vector book (referred to as VQ code book below) of a standard speaker in speech recognition is adapted to an unknown input speaker, and its effectiveness is confirmed by the evaluation by large vocabulary discrete word recognition. There is. In the SPLIT type speech recognition, the standard pattern is VQ.
The vector is quantized by a codebook and stored in the dictionary as a code label string. Conventionally, a clustering method of a phonological feature parameter space has been used as a standard pattern generation method of this kind, and for example, a document “Tsubomura, Hakoda, Furui,“ Word Speech Recognition Using Phonological Standard Pattern by Clustering ”, Acoustical Society of Japan Material for Speech Research S80-61
There is a method described in "December 15, 1980", and Fig. 2 shows this method in a block diagram.

In the figure, (1) is an input end, (2) is an input audio signal,
(3) is an acoustic analysis means, (4) is a feature vector time series,
(5) is a changeover switch, (6) is clustering means, (7) is centroid vector, (8) is codebook memory, (9) is code vector, (10) is quantization means, and (11) is label. The column, (12), is a dictionary memory.

Next, the operation will be described.

First, in order to create a VQ codebook, the changeover switch (5) is set to the a side, and the voice for creating the VQ codebook is input from the input end (1). The input audio signal (2) of M frames, which is input from the input terminal (1), is acoustically analyzed by the acoustic analysis means (3), and the feature vector time series (4) {X (1), X
(2), ..., X (M)}. Feature vector time series (4) {X which is the output of the acoustic analysis means (3)
(1), X (2), ..., X (M)} are input to the clustering means (6). The clustering means (6) clusters all the input feature vectors of the feature vector time series (4) and outputs a centroid vector (7). The centroid vector (7) is stored in the codebook memory (8) as a VQ codebook.

After the VQ codebook is created, create a dictionary of standard patterns for recognition. The changeover switch (5) is set to the b side, and the standard pattern voice is input from the input end (1). The input voice signal (2) of the M'frame input from the input terminal (1) is acoustically analyzed by the acoustic analysis means (3), and the feature vector time series (4) {X (1), X (2), ... X (M ')}. The feature vector time series (4) {X (1), X (2), ..., X (M ')} which is the output of the acoustic analysis means (3) is input to the quantization means (10). The quantizing means (10) uses a feature vector time series (4) {X (1), X (2), ..., X.
The code vector (9) having the maximum likelihood for the feature vector X (m) in (M ′)} is searched from the codebook memory (8), and the label L of the code having the maximum likelihood is searched.
(M) is obtained. This is performed for m = 1,2, ..., M 'to generate a label string (11) {L (1), L (2), ..., L (M')} and register it in the dictionary memory (12). To do.

[Problems to be Solved by the Invention] In the speaker adaptation method using the standard pattern described above, when the standard pattern according to the conventional technique described above is used, there are the following problems. That is, since the standard pattern is created from the voice of one standard speaker, the vocalization habits of the standard speaker and the individuality of the articulation due to the habit are preserved as they are, and therefore the vocalization habits of unknown input speakers are preserved. There is a problem that the articulation pattern may differ from the individuality due to habits, and the effect of adaptation may differ between standard speakers, and the recognition rate after adaptation may not be the same as that of specific speakers. Although there is a multi-template method for dealing with the individuality of various articulatory patterns by using multiple standard speakers, there is a problem that the amount of processing at the time of recognition increases. There is also a method of averaging the voices of a plurality of standard speakers, but the phonological features are blurred and the recognition rate is reduced.

The present invention has been made to solve the above problems, and an object of the present invention is to create a standard pattern having a small individual difference in articulation mode.

[Means for Solving the Problem] A standard pattern creating method according to the present invention includes a time axis normalizing means for normalizing a feature vector time series of voices of the same category uttered by a plurality of standard speakers with respect to a time axis. Dimension extension means for converting a plurality of time axis normalized feature vector time series output from the time axis normalization means into one dimension extension feature vector time series with an expanded number of dimensions, and dimensions output from this dimension extension means. Input extended feature vector time series,
After performing clustering on the dimension extension feature vector, a dimension extension clustering unit that outputs the centroid vector of each cluster, a memory that stores the centroid vector as a dimension extension VQ codebook, and the dimension extension VQ codebook are stored. Using the dimensional expansion feature vector time series, the dimensional expansion quantization means for vector quantizing and outputting the label string, the dictionary memory storing the label string as a dictionary, and the dimensional expansion feature of each code of the dimensional expansion VQ codebook It is characterized by comprising dimension reduction means for returning the dimension number of a vector to the original dimension number of a feature vector, and a codebook memory for storing the dimension-reduced feature vector.

[Operation] The time axis normalizing means is a time axis based on the feature vector time series of an arbitrary standard speaker with respect to the feature vector time series of the same category speech uttered by a plurality of standard speakers The normalization processing is performed and the time-axis normalized feature vector time series is output. The dimension extension means outputs one dimension extension feature vector time series using a plurality of time axis normalized feature vector time series in the same category. The dimension extension clustering means clusters all the dimension extension feature vectors of the input time series of dimension extension feature vectors, obtains a dimension extension centroid vector for each cluster, and stores it in the dimension extension codebook memory. The dimension extension quantization means refers to the dimension extension codebook, performs vector quantization of the dimension extension feature vector time series, and stores the label string in the dictionary memory. The dimension reduction means reduces the dimension number of the dimension extension feature vector of each code on the dimension extension codebook memory to the dimension number of the original feature vector, and stores the dimension reduction feature vector in the codebook memory.

[Embodiment] FIG. 1 is a block diagram showing a configuration of a standard pattern creating method according to an embodiment of the present invention. In FIG. 1, components corresponding to those shown in FIG. 2 are designated by the same reference numerals, and their description will be omitted.

In FIG. 1, (13) is a feature vector time series (4) of a plurality of standard speakers of the same category, which is an output of the acoustic analysis means (3), and a feature vector time series of an arbitrary standard speaker. Time-axis normalization means that performs time-axis normalization processing by DTW based on the sequence and outputs the time-axis normalized feature vector time series (14), (15) is the time of the same category speech of multiple standard speakers Axis normalization feature vector time series (14)
A dimensional expansion means (17) for converting into one dimensional expanded feature vector time series (16) receives the dimensional expanded feature vector time series and performs clustering processing to obtain the dimensional expanded centroid vector (18) of each cluster. Dimension extension clustering means for outputting, (19) dimension extension codebook memory for storing dimension extension centroid vector (18) as dimension extension VQ codebook, (21) dimension extension on dimension extension codebook memory (19). The dimensional expansion feature vector time series (16) is vector-quantized using the VQ codebook, and the dimensional expansion quantization means (22) for outputting the label sequence (11) is stored in the dimensional expansion codebook memory (19). Dimension reduction feature vector (23) by reducing the number of dimensions of the dimension extension code vector of each code in the dimension extension VQ codebook.
Is a dimension reduction means for storing in the codebook memory (8).

The same category speech uttered by a plurality of (N) standard speakers is subjected to N groups of feature vector time series (4) {X ⁽ n ⁾ (1), X ⁽ n ⁾ (2), by the acoustic analysis means (3). …, X ⁽ n ⁾ (Mn)} (n
= 1,2, ..., N). Here, Mn is the sequence number of the feature vector time series of the speaker n. Time axis normalization means (13)
Is the N-group feature vector time series (4) {X ⁽ n ⁾ (1), X ⁽ n
⁾ (2), ..., X ⁽ n ⁾ (Mn)} (n = 1,2, ..., N) is input and the feature vector time series of the standard speaker n {X ⁽ n ⁾ (1), X ⁽ n
⁾ (2), ..., X ⁽ n ⁾ (Mn)} is used as a reference, and DTW (Dyn (Dyn)
amic Time Warping), the time series normalized feature vector time series of N groups (14) {X ′ ⁽ n ⁾ (1), X ′ ⁽ n
⁾ (2), ..., X ' ⁽ n ⁾ (M)} (n = 1,2, ..., N) is output. Here, M is equal to the number Mn of series of feature vector time series of the standard speaker n. The dimension expanding means (15) is a time series normalized feature vector time series (14) {X ' ⁽ n ⁾ (1), X' ⁽ n ^{) of} N groups.
Using (2), ..., X ' ⁽ n ⁾ (M)} (n = 1,2, ..., N),
Dimensional extended feature vector time series (16) {Y (1), Y
(2), ..., Y (M)} is obtained as follows.

Let yk be the k-dimensional element of the dimension extension feature vector Y (m), and xn, j be the j-dimensional element of the time-axis normalized feature vector X ' ⁽ n ⁾ (m) of the speaker _{n. n, j} (n = 1,2, ..., N, j = 1,2, ..., J) However, k = f (n, j).

Here, N is the number of standard speakers, and J is the number of dimensions of the feature vector. The function f is defined by n = 1,2, ..., N, j = 1,2, ..., J,
All combinations of n and j and k are functions that have a one-to-one correspondence. For example, a function such as f (n, j) = (n-1) * J + j, and the dimension extension processing in this case is shown in FIG.

By obtaining this for m = 1, 2, ..., M, the time-axis normalized feature vector time series (14) {X ′ ⁽ n ⁾ (1), X ′ of N-group standard speakers in the same category. ⁽ n ⁾ (2),…, X ′ ⁽ n
⁾ (M)} (n = 1,2, ..., N) is a one-dimensional extended feature vector time series (16) {Y (1), Y () while preserving the individuality of the articulatory aspect of each speaker. 2), ..., Y (M)}.

Dimensional expansion vector time series obtained in this way (16)
Create a VQ codebook and dictionary using {Y (1), Y (2), ..., Y (M)}. First, the changeover switch (5) is set to the a side to create a VQ codebook. A dimension extension clustering means (17) is a dimension extension feature vector time series (16) {Y (1), Y (2), ..., Y (M) with the number of sequences M generated from the voice for VQ codebook creation. } As an input, clustering is performed, and the dimension expansion feature vector (18) Zi | i = 1,2, ..., I} of the centroid of each cluster (I
Stores the number of clusters) in the dimension extension codebook memory (19). The dimension reduction means (22) receives the dimension extension code vector (20) Zi of each code of the dimension extension VQ codebook stored in the dimension extension codebook memory (19) as an input, and the dimension reduction feature vector for the standard speaker n. Z ⁽ n ⁾ i
Is calculated as follows.

Let x′j be the j-dimensional element of the dimension reduction feature vector Z ⁽ n ⁾ i,
If the k-dimensional element of the dimension extension code vector Zi is Zk, then z'j = zk (j = 1,2, ..., J) However, k is used in dimension extension because j and n are given. Function f (in the above example, k = (n-1) * J + j
Then, the processing in this case is shown uniquely in FIG.

This is obtained for i = 1, 2, ..., I, and each dimension reduction feature vector (23) Z ⁽ n ⁾ i is stored in the codebook memory (8).

Next, change the selector switch (5) b to create a dictionary.
Turn it to the side. The dimension extension quantizing means (21) is a dimension extension feature vector time series (16) {Y (1), Y (2), ..., Y (M ')} having the number of sequences M'created from the standard pattern speech. For each dimensional expanded feature vector Y (m) of, the code having the maximum likelihood is searched from the dimensional expanded VQ codebook in the dimensional expanded codebook memory (19) to obtain the label L (m). By obtaining this for m = 1, 2, ..., M ′, the label sequence (1
1) Generate {L (1), L (2), ..., L (M ')}.
The generated label string is stored in the dictionary memory (12).

The dictionary thus obtained is vector-quantized with respect to one dimension-extended feature vector time series generated by multiplexing the feature vector time series of a plurality of standard speaker voices by dimension extension, and creates a label string. As a result, the individuality of the articulation pattern due to the speaking habits and habits of each speaker is smoothed, and the individuality of the articulation pattern is reduced. Therefore, there is no difference in the adaptation effect due to individual differences in articulatory patterns of the standard speaker and the input speaker, which was a problem in speaker adaptation processing, and the recognition rate after adaptation is significantly improved compared to the conventional method. To do.

In the present invention, the implementation method may be dedicated hardware or software implementation in a general-purpose computer or a signal processor.

EFFECTS OF THE INVENTION As described above, according to the present invention, the time axis normalizing means for normalizing the time axis of the feature vector time series of the voices of the same category of a plurality of standard speakers, and the output of this time axis normalizing means Dimension expanding means for expanding a time series normalized feature vector time series that is the same category speech of a plurality of standard speakers into one dimension expanded feature vector time series,
A clustering means for performing clustering using the time series of the dimension expansion feature vector output from the dimension expansion means as an input, and outputting the dimension expansion feature vector of the centroid of each cluster, and the dimension expansion feature vector for the dimension expansion VQ codebook. , A vector quantization means for quantizing a dimension extension feature vector time series using this dimension extension VQ codebook, and converting it into a label sequence of a corresponding code, and a memory storing this code label sequence as a dictionary. And a dimension extension feature vector of each code of the dimension extension VQ codebook, and a dimension reduction means for reducing the number of dimensions of the dimension extension feature vector of each code, and a memory for storing the dimension reduction feature vector with the reduced number of dimensions. Since it is composed of and, the same category of voices of multiple standard speakers is analyzed by the phonological features of each speaker. Can be converted into a one-dimensional extended feature vector time series without any effort, and VQ codebooks and dictionaries created by performing clustering and vector quantization on them are standard patterns with less individuality in articulation. This has the effect of significantly improving the recognition rate when speaker adaptation is performed using this.
Also, it goes without saying that this method can be applied not only when using multiple standard speakers but also when creating a standard pattern using the same Kadekori voice produced by one speaker multiple times. A stable standard pattern that is not affected by vocalization fluctuations can be obtained, which has the effect of improving the recognition rate.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a standard pattern creating device according to an embodiment of the present invention, FIG. 2 is a block diagram showing a configuration of a conventional standard pattern creating device, and FIG. 3 is a process of dimension expanding means. FIG. 4 is a diagram showing an example, and FIG. 4 is a diagram showing a processing example of the dimension reducing means. Reference numeral (8) in the figure ... Dimension expansion codebook memory, (1
3) …… Time axis normalization means, (15) …… Dimension extension means,
(17) Dimension expansion clustering means, (19) Dimension expansion codebook memory, (21) Dimension expansion quantization means, (22) Dimension reduction means. The same reference numerals in the drawings indicate the same or corresponding parts.

Claims (1)

[Claims] 1. A code label obtained by creating a vector quantization codebook of a standard speaker's voice in voice recognition and vector-quantizing the standard speaker's voice using this vector quantization codebook. In a standard pattern creating method for storing a sequence as a dictionary, a time axis normalizing means for normalizing the feature vector time series of the same category speech uttered by a plurality of standard speakers with respect to the time axis, and the time axis normalizing means. Dimension expansion that multiplexes the feature vectors that exist for the number of standard speakers at the same point in time to multiple feature vector time series normalized by Means and the dimensional extension feature vector in the dimensional extension feature vector time series, and outputs the centroid dimensional extension feature vector of each cluster. Clustering means, a dimension extension codebook memory for storing the dimension extension feature vector output from this clustering means as a dimension extension vector quantization codebook, and a dimension extension vector quantization codebook in this codebook memory. , A quantizing means for vector-quantizing the time series of the dimension expansion feature vector output from the dimension expanding means and outputting a code label sequence, a dictionary memory for storing the code label sequence as a dictionary, and the dimension expansion vector quantization Dimension reduction means for reducing the number of dimensions to the number before extension by extracting the feature vector corresponding to each speaker of the plurality of standard speakers from the dimension extension feature vector of each code of the code book, and the dimension reduction. A codebook memory that stores the feature vector as a vector quantization codebook Standard pattern creating system configured.