JPS61261799A - Code book preparation for unspecified speaker - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to techniques applied to waveform encoding systems for audio signals, etc., and linear predictive analysis/synthesis systems, and relates to a clustering method for pattern recognition of audio and images. It can be applied to

In general, the clustering method for extracting representative vectors (called code vectors) from the characteristics of audio and images is called vector quantization, and recently, waveform encoding methods, linear predictive analysis, It is applied to synthesis systems, code transmission systems, etc.

An efficient way to create code vectors is to cluster them according to the spatial distribution of training sample data, but no definitive method has been established yet, and quantization is usually done through a large number of repeated operations. A typical algorithm is one that minimizes the average distortion (error between the code vector and the learning sample) when

Conventionally, creating code vectors requires a huge amount of calculation, and in particular, as the number of code vectors and training samples increases, the amount of calculation increases proportionally. Even if there was a chance, the drawback was that it would be difficult to create in reality.

One Purpose The present invention has been made in view of the above-mentioned circumstances.
In particular, codeb is able to perform fast and accurate processing for repeated operations on the huge amount of learning sample data used in vector quantization.

This was done for the purpose of efficiently creating a set of code vectors.

Structure: In order to achieve the above object, the present invention extracts a representative vector (referred to as a code vector) from a set of feature vectors (referred to as a learning sample) of audio, images, etc.

Also, a collection of code vectors is called a codebook. )
In the vector quantization method, which is created based on a clustering method, code vectors are created by vector quantization for each speaker among multiple speakers, and then the number of learning samples belonging to those code vectors is registered, This method is characterized by weighting the code vectors of all multiple speakers by the number of learning samples and vector quantizing them again to quickly create code vectors for unspecified speakers.

EMBODIMENT OF THE INVENTION Hereinafter, the present invention will be explained based on an example.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, and in the figure, ■ indicates a signal input terminal.

2 is a feature analysis section, 3 is a codebook (a collection of code vectors) storage section, 4 is a vector quantization section, 5 is a standard pattern storage section 56 is a recognition processing section, and 7 is a recognition result output terminal.

Generally, image signal input, audio signal input, etc. can be considered as signal input, but only audio signals will be explained here.

The signal input at the input terminal 1 is subjected to analysis processing such as a group of band-pass filters and r, pc analysis in the characteristic analysis section 2, and is converted into characteristic parameters such as power spectrum and LPG parameters. . Using a set of these feature parameters as a learning sample, representative vectors (referred to as code vectors) are created in advance by a clustering method (vector quantization method) described later and stored in the codebook 3. A collection of these code vectors is called a codebook.

A certain signal input is vector quantized into one of the code vectors in the code book 3 by the vector quantizer 4, and the code (the number of the yaw 1-' vector) is registered for each frame. In the standard pattern storage section 5, those corresponding to dictionary patterns are expressed by codes in the codebook 3. In this way, the unknown input cover pattern vector quantized by the vector quantization unit 4 and the standard pattern storage unit 5
The recognition processing unit 6 performs pattern matching with the standard pattern of
Then, the pattern most similar to the unknown input cover turn is selected from the standard pattern storage section 5 and outputted at the output terminal 7.

FIG. 2 shows a flowchart based on a typical vector quantization algorithm. This algorithm minimizes the average distortion that occurs during quantization, and is one of the most powerful algorithms. In the figure, 8 is the algorithm start terminal.

9 is an initialization section, and 10 is a learning sample reading section.

ll is a clustering part, and 12 is an average distortion calculation part.

13 is a convergence determination unit, and 14 is an average distortion replacement unit.

15 is a codebook determining section, 16 is a codebook storage section, and 17 is an algorithm termination terminal.

First, the initialization unit 9 uses the average distortion D-1 for determining convergence.
(set as D-1-■), setting of quantization level N,
Initial vector Y1. )'z, . . . , the setting of the adhesive 4 and the convergence determination threshold ε are performed. As the quantization level N, i
11, 128.256 for speech recognition devices for specific speakers
It is said that a level such as 256, 512 etc. is recommended for non-specific speakers. Initial vector (i
t l , (i=].

2,...N), the learning sample X1° cross 2. It is normal to take N vectors that are as different from each other as possible from among <n. Otherwise,
There is a risk that clustering will not be performed correctly and converge locally. Also, the number of learning samples n is usually
It is preferable to set the quantization level N to 10 times the borrowed background. As the dark value for convergence determination, for example, ε=O,'OOI is selected. The reading unit 10 reads the learning sample X1. X2. ..., read the intersection n into the buffer memory, and for each sample intersection,
Initial code vector - y1. 'tz,...,9,
Calculate the distance 11 moxibustion j - 11 from 4. At this time, select 9, which has the smallest distance, and consider that Xi belongs to Tadasu. In this way, all the learning samples "It Ai 2"
+...

Clustering for M n is completed. 9° each (
i =], 2, ..., N) as a code vector is called 01, then all the learning samples M4 and 9. Calculate the distance l1itl (referred to as i quantification distortion) to all clusters Ci (i=1,2°...,N
) is calculated as the average distortion Do by the average distortion calculation unit 12. Next, the (initial) distortion value D−,
The relative change (D-1-Do)/Do between and Do is calculated, and if it is smaller than the threshold ε predetermined by the initialization unit 9, then fy+l at this time.

(i=], 2..., N) is determined as the final codebook by the codebook determining unit 15, and each cluster O1
After registering in the memory unit 16 together with the number of samples 81 belonging to the sample, the process ends. However, the above is the total amount from the clustering unit 11 to the convergence determination unit 13! Even if the initial vector (9,) of the process ill initialization unit 9 is selected appropriately, it will take several times to converge.
- Requires repetition. Therefore, the replacement part 14
Then, the value of DO is replaced with D-+, and the process returns to the clustering unit 11 again to repeat clustering and calculation of DO until the conditions of the convergence determination unit 13 are satisfied.

FIG. 3 is a diagram showing the learning samples (XJl, (j=1.2..., n), each cluster Ci, etc. at the time when the codebook has been created, and each vector is generally P (P ≧2) dimension, but the figure shows the case where p=2.

FIG. 4 shows a flowchart 1 based on the algorithm of the speaker-independent codebook creation method according to the present invention, and the basic configuration is the same as that in FIG. 2. In the initialization section 19, the average strain l)-τ is calculated as in the initialization section 9 of FIG.
-■, and ε are set, and the bell is N1. The initial vector is a code vector for a particular speaker 1, as will be described later in FIG. For unspecified speakers, the N slag is 2-4';'i for specific speakers;
In other words, it is said that approximately N''=2N is good.Reading section 2
0, speaker m (m=12.-, M; M is the number of speakers) codebook fyTl and its number f371 were created in advance in codebook 16 by the method described in Fig. 2. Load. A clustering section 21 performs clustering, and a calculation section 22 calculates an average distortion D01. Here, the calculation of the DO-tei is performed as follows. That is, the initial vector (shi1) and the learning sample (yTl
, (j=l, 2....5 N; m=2.3.
..., M), the nearest si j (i=], 2..... for each yW.

The distance to N) multiplied by 3+i2 as a weight is defined as distortion, and all clusters Ci'' (i =], 2°...
・, N”). Expressed in the formula, however,
It is n-N-M.

The relative distortion is calculated by the convergence determination unit 23 as described in -, and if the change is smaller than ε, the codebook at this point is calculated as (夛71. (i = 1.2°..., N1
) is determined by the codebook determining unit 25, and is registered in the memory of the codebook storage unit 26 as a codebook for unspecified speakers, and then the process ends. If the determination by the convergence determining section 23 is negative, the clustering by the clustering section 21 is restarted, which is the same as described in FIG. 2.

Figure 5 shows speaker 1. ...1m, ...9M code book sp1. . . , SPm, and SPH. In the diagram, T, . Represents the number of training samples vector quantized into the code vector.

FIG. 6 is a diagram showing a code vector for unspecified speakers created with the unspecified code book 26 of FIG. 4,
In the figure, yT' and ST are the code vectors of speaker m, respectively,
The number of learning samples (weight) belonging to it, 9, and ST are the code vector for unspecified speakers and the number of learning samples belonging to it, respectively. In the initialization unit 19 of FIG. 4, when setting the initial vector, Coat for identifying speaker 1 [
・Although we used M, it goes without saying that speaker l may be any speaker in M.

Furthermore, in the algorithm shown in Fig. 4, code vectors for two people can be created by vector quantizing the learning sample of speaker 2 according to the algorithm described in Fig. 2 based on the initial vector of talk -i+. Similarly, after repeating this operation to create code vectors for speakers 1 to M (1<m<M), the next speaker (
m+1) training samples may be vector quantized.

By performing such an operation, it is possible to prevent code vectors for unspecified speakers from converging locally and to create them in a well-balanced manner and at high speed.

Next, a statistical processing method that takes into account the variance of learning samples within each cluster will be described.

To find the variance and covariance of the learning samples in each cluster Ci in FIG. 3, the covariance matrix Σci can be defined as follows.

Σci= (ty jk), j, k= 1.
2. -, P.

Here, P is the number of parameter dimensions, and σjk is the N, k) component of the covariance matrix, which can be expressed as follows. x1□j is the j component of the first learning sample in cluster C4, and y, 4 is the J component of the code vector of C4.

The above covariance matrix Σci is the Mahalanobis distance d (X+ - Y+ ) - (X+, )', + )
Σ昌 (Intersection 1-Ten) It is often used when calculating T. Here, -1 represents the inverse matrix and T represents the transposition. When calculating Σci or Σε1, the amount of calculation increases as the number of learning samples increases. do.

Therefore, using the Mahalanobis distance as described above from the beginning as a distance measure between vectors during vector quantization requires a huge amount of time and becomes an unrealistic calculation. To solve this problem, until the codebook is determined by the codebook determination unit 15 in FIG. If we calculate the above equation (1) after convergence, we will get convergence 8
Ability to perform calculations efficiently.

In addition, the covariance matrix Σc1 in the cluster C4 to which the code vector for the general speaker in the non-specific code 1-'' book 2G in FIG. By multiplying the weight by 87, it can be defined as follows.

ΣcY−(ty +;), j, k= 1.
2. −, p... (2) However, if the covariance matrix is defined as in equation (2) above, the amount of calculation required for statistical processing of learning samples for non-specific speakers can be greatly reduced. This can be reduced to a practically possible process.

Effects As is clear from the above explanation, the present invention reduces the amount of repeated calculations when creating a codebook for an unspecified speaker with a high quantization level and a large number of learning samples. Since the initial vector is selected in a well-balanced manner to avoid local convergence, the amount of calculation can be significantly reduced compared to conventional methods, making it possible to perform accurate and realistic processing.

[Brief explanation of the drawing]

Fig. 1 is an electrical block diagram for explaining an embodiment of the present invention, Fig. 2 is a flowchart showing the vector quantization algorithm, and Fig. 3 shows the relationship between learning samples and clusters. Figure 4 is a flowchart showing the algorithm for creating a codebook for non-specific speakers.
The figure shows the codebook creation result, and FIG. 6 is a diagram showing the coat vector for unspecified speakers. DESCRIPTION OF SYMBOLS 1... Signal input terminal, 2... Feature analysis section, 3... Code book, 4... Vector quantization section, 5... Standard pattern. 6... Recognition processing section, 7... Result output section, 8.18.
...Algorithm start terminal, 9.19...Initialization section,
10゜20...Learning sample reading section, 11.21
. . . Clustering section, 12.22 . . . Average distortion calculation section. 13.23... Convergence determination section, 14.24... Average distortion replacement section, 15.25... Codebook determination section. 16.26...Codebook storage section, 17.27...
・Algorithm end terminal. Figure 3 Figure 4 Bunro Nu old first rib Ishiyoshi Ryochi mn code book (9T) and number of de S T 4 calculation Yosakami cluster 1 Nonghira 7 distortion Ninogi Shisan 22 Fig. 5

Claims (5)

[Claims] (1) In the vector quantization method, which creates representative vectors from a set of feature vectors of voices, images, etc. based on a clustering method, vector quantization is performed for each speaker among multiple speakers to create code vectors. After creating , the number of training samples belonging to those code vectors is registered, and the code vectors for all multiple speakers are weighted by the number of training samples and vector quantized again to create a code vector for unspecified speakers. A method for creating a codebook for unspecified speakers, characterized by the following. (2) Claim No. 1 characterized in that learning samples of other speakers are vector quantized using a code vector of a specific speaker among a plurality of speakers as an initial vector. ) The method for creating a codebook for non-specific speakers as described in section 2. (3) When creating a codebook for the (m+1)th speaker or later, vector quantization is performed using the codebook created from speakers 1 to m as an initial vector. 1) Method for creating a codebook for non-specific speakers as described in section 1). (4) Calculate the Mahalanobis distance from the covariance matrix using the code vectors for each speaker that belong to the created code vectors for unspecified speakers, calculate the statistical inter-code vector distance, and then calculate the vector quantum A method for creating a codebook for an unspecified speaker according to claim (1), characterized in that: (5) Use absolute value distance or Euclidean distance, which requires less calculation, until the code vector converges as a distance measure between vectors during vector quantization, and then switch to Mahalanobis distance after convergence to create a codebook. A method for creating a codebook for unspecified speakers according to claim (4).