JPH0339800A - Optimum vector code word search device - Google Patents

Abstract

PURPOSE:To make a fast survey by converting a multi-dimensional input vector into a multi-dimensional degenerative input vector which is smaller in the number of dimensions and calculating the distance between the degenerative- dimensional input vector and each intermediate vector in an intermediate code book. CONSTITUTION:A dimension degenerating means 15 converts the L-dimensional input vector into the J-dimensional degenerative-dimensional input vector which is reduced in the number of dimensions, a calculating means 16 for a distance in degenerative-dimensions calculates distances between the degenerative- dimensional input vector and intermediate vectors in the intermediate code book 10, and an optimum intermediate vector selecting means 6 selects the intermediate vector which has the shortest distance. Then a distance calculating means 12 calculates the distance between the input vector and a vector code word specified by a code word number table 8 corresponding to the selected intermediate vector and an optimum vector code word selecting means 13 selects and outputs the vector code word which is given the shortest distance. Consequently, the arithmetic quantity is reduced and the fast survey is made.

Description

[Detailed Description of the Invention] [Industrial Application Field] This invention compresses information such as audio signals and image signals using a codebook consisting of a finite plurality (N) of plurality (L)-dimensional vector codewords. The present invention relates to an apparatus for searching for an optimal vector codeword.

[Prior Art] Conventionally, as a method for compressing input vectors such as parameters and graphic patterns representing the frequency spectrum characteristics of speech into a small amount of information, a codebook consisting of a finite plurality (N) of vector codewords is prepared in advance. There is a vector quantization method in which the vector code word closest to the input vector is selected from among the vector code words, and the input vector is expressed using only that code number.

In the optimal vector codeword search device using this vector quantization method, the number of calculation stars used to search for the optimal vector codeword that best represents the input vector increases in proportion to the number of codewords in the codebook, so real-time In applications that require processing, it is necessary to speed up the search processing.

A typical example of an optimal vector codeword search device that speeds up conventional search processing is given in the Acoustical Society of Japan Lecture Proceedings 2-
7-12 (October 1982) Kiyohiro Kano 1 Word speech recognition by vector quantization of input speech", M-3 for the number of vector code words in the codebook N-512
By providing two intermediate vectors, associating in advance which intermediate vector each vector codeword in the codebook is close to, and calculating the distance only between the intermediate vector closest to the input vector and the corresponding vector codeword. , which reduces exploration compared to total exploration.

FIG. 2 is a block diagram showing an example of the configuration of a conventional optimal vector codeword search device.
It has an input terminal (2) into which is inputted, and the input terminal (
2) is connected to distance calculation means (4) for calculating the distance between the input vector and each intermediate vector in the intermediate codebook (3) on a distance evaluation scale given in advance.

Further, the distance calculation means (4) is connected to the intermediate codebook (3) via a changeover switch (5), and the changeover switch (5) sequentially selects M intermediate vectors of the intermediate codebook (3). It is supposed to connect.

Further, on the output side of the distance calculation means (4), there is provided optimal intermediate vector selection means (6) for selecting the intermediate vector that gives the smallest distance among the distances calculated by the distance calculation means (4).
) are connected, and the optimal intermediate vector selection means (6)
On the output side, M code number tables (8a) corresponding to the intermediate vectors selected by the optimal intermediate vector selection means (6) are provided.
A code number table (8) is connected through a changeover switch (7) which can be switched to select one of the following.

And the code number table (8) is M code number tables (8a)
and a changeover switch (9) for sequentially reading out the code numbers in each code number table (8a).

Further, a changeover switch (11) for selecting a vector codeword in the codebook (10) is connected to the changeover switch (7), and a changeover switch (11) for selecting a vector codeword in the codebook (10) is connected to the changeover switch (11). It is also connected to distance calculation means (12) for calculating the distance on a distance evaluation scale between the vector code word selected by the changeover switch (11) and the input vector (100) given in advance.

Further, on the output side of the distance calculation means (12), the vector code word calculated by the distance calculation means (12) and the input vector (
100) Optimal vector code word selection means (13) for selecting the vector code word that gives the smallest distance among the distances between
is connected to the optimal vector code word selection means (13
) is connected to the output terminal (14).

In addition, in the document "Word speech recognition by vector quantization of input speech", since the purpose is to perform speech recognition, the WLR distance measure and the PWLR distance measure are often used as distance measures between feature parameters of the speech spectrum. ing.

Hereinafter, the distance on the distance evaluation scale given in advance will be simply referred to as distance.

Next, the operation will be explained.

The characteristic parameters of the audio spectrum are inputted from the input terminal (2) as an input vector (100) to the distance calculation means (4) and (12).

The distance calculation means (4) is a changeover switch (5).
calculates the distance between the intermediate vectors of the intermediate codebook (3) and the input vector (100) that are sequentially input, and outputs the calculated distance to the optimal intermediate vector selection means (6).

The optimal intermediate vector selection means (6) comprises a distance calculation means (4).
) selects the intermediate vector that gives the minimum distance among the input distances, and switches to select one code number table (8a) from the code number tables (8) corresponding to that intermediate vector. Switch the switch (7).

Then, the changeover switch (11) is set to the code number table (8
The vector codeword in the codebook (10) indicated by the code number sequentially read out from a) by the changeover switch (9) is selected and output to the distance calculation means (12), and the distance calculation means (12) Vector (100) and changeover switch (
11), and outputs the calculated distance to the vector code word inputted through the vector code word selection means (13).

Furthermore, the optimal vector code word selection means (13) selects the vector code word that gives the minimum distance among the input distances, and outputs its code number via the output terminal (14).

Therefore, compared to the exhaustive search method, which searches for the optimal vector codeword by calculating the distance between the input vector and all vector codewords, the search can be performed N/(M+K) times faster.

[Problems to be Solved by the Invention] Since the conventional optimal vector codeword search device has the above configuration, it is difficult to perform optimal intermediate vector search processing for roughly classifying input vectors using intermediate vectors of Ml. , for example, in the case of M=, it occupies half of the total exploration processing, and there is a problem that the load is large.

Therefore, the present invention was made to solve this problem, and an object of the present invention is to obtain an optimal vector codeword search device that can perform a high-speed search even in the case of M-.

[Means for Solving the Problem] The optimal vector codeword search device according to the present invention converts a plurality (L)-dimensional input vector into a plurality (J)-dimensional input vector with a reduced number of dimensions.
)-dimensional reduced-dimensional input vector; an intermediate codebook comprising a plurality (M) of plural (J)-dimensional intermediate vectors smaller than the finite plurality (N); and a reduced-dimensional input vector; Optimization that selects the intermediate vector that gives the minimum distance among the distances calculated by the degenerate second-order distance calculation means that calculates the distance between each intermediate vector in the intermediate codebook and the degenerate dimension distance calculation means. A plurality of (K) codewords that are close in distance between the intermediate vector selection means and each intermediate vector selected in advance from among the vector codewords in the codebook for each intermediate vector and are smaller than the finite plurality (N). Each vector code word and input vector specified by a plurality (M) of code word number tables in which code addresses of sets are listed and a code word number table corresponding to the intermediate vector selected by the optimal intermediate vector selection means. and an optimal vector code word selection means that selects and outputs the vector code word that gives the smallest distance among the distances calculated by the distance calculation means. .

[Operation] The optimal vector codeword search device according to the present invention converts a multi-dimensional (L)-dimensional input vector into a multi-dimensional (J)-dimensional reduced-dimensional input vector with a reduced number of dimensions using a dimension reduction means, The distance between the degenerate dimensional input vector and each intermediate vector in the intermediate No. 71 book is calculated by the degenerate dimension short distance calculation means, and the minimum distance is given among the distances calculated by the degenerate second dimension distance calculation means. The intermediate vector selected by the optimal intermediate vector selection means is selected by the optimal intermediate vector selection means, and the distance calculation means calculates the distance between each vector code word specified by the code word number table corresponding to the intermediate vector selected by the optimal intermediate vector selection means and the input vector. The optimal vector code word selection means selects and outputs the vector code word that gives the smallest distance among the distances calculated by the distance calculation means.

[Example code] An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an example of the configuration of an optimal vector codeword search device.
) has an input terminal (2) into which a plurality (L)-dimensional input vector (100) is input; A dimension reduction means (15) is connected for converting into a reduced dimension input vector of reduced multiple (J) dimensions.

Then, on the output side of the dimension reduction means (15), the reduced dimension input vector and a plurality (M
) A distance between each intermediate vector in the intermediate codebook (3) consisting of a plurality of (J)-dimensional intermediate vectors and a pre-given vector, 71' (degenerate dimensional distance for calculating the distance on the dE scale) Calculating means (16) are connected.

Further, the degenerate dimensional distance calculation means (16) is connected to the intermediate codebook (3) via the changeover switch (5),
The changeover switch (5) is configured to sequentially connect M intermediate vectors of the intermediate codebook (3).

Furthermore, on the output side of the degenerate second-order distance calculation means (16),
Optimal intermediate vector selection means (6) is connected to select the intermediate vector that gives the smallest distance among the distances calculated by the degenerate second-order distance calculation means (16). ) is provided with a changeover switch (7) that can be switched to select one of the M code number tables (8a) corresponding to the intermediate vector selected by the optimal intermediate vector selection means (6). A code number table (8) is connected thereto.

And the code number table (8) is M code number tables (8a)
and a changeover switch (9) for sequentially reading out the code numbers in each code number table (8a).

Further, a changeover switch (11) for selecting a vector codeword in the codebook (lO) is connected to the changeover switch (7), and a changeover switch (11) for selecting a vector codeword in the codebook (lO) is connected to the changeover switch (11). Also connected is a distance calculation means (12) for calculating the distance on a distance evaluation scale between the vector code word selected by the changeover switch (11) and the input vector (100) given in advance.

Further, on the output side of the distance z-1 calculating means (12), a vector code word that gives the smallest distance among the distances between the vector code word calculated by the distance calculating means (12) and the input vector (100) is provided. Optimal vector codeword selection means (1
3) is connected, and the optimal vector codeword selection means (
An output terminal (14) is connected to the output side of 13).

Next, the operation of this embodiment will be explained.

An L-dimensional input vector (100) is input from an input terminal (2) to a dimension reduction means (15) and a distance calculation means (12).

Then, the dimensional reduction means (15) converts the L-dimensional input vector (100) into a J-dimensional (J<L) reduced-dimensional input vector and outputs it to the reduced-dimensional distance calculation means (16).

In addition, in the dimension reduction method in the dimension reduction means (15), as a method with the least amount of calculation, J dimensions are selected in descending order of influence on a distance evaluation scale given in advance among L dimensions, and the input There is a method of extracting five components from the L components of a vector and creating a J-dimensional degenerate dimensional input vector.

For example, if the distance evaluation measure is Euclidean distance, calculate the fraction for each dimension for the input vector set,
It is sufficient to select J dimensions in descending order of their values. Also, L
It is also possible to calculate and use a J-dimensional principal component vector from the dimensional intermediate vector.

Then, the degenerate dimensional distance calculation means (step 6) calculates the intermediate codebook (3
) and the degenerate dimensional input vector, and output the calculated distance to the optimal intermediate vector selection means (6).

The optimal intermediate vector selection means (6) selects the intermediate vector that gives the minimum distance among the distances input by the degenerate second-elastic distance calculation means (16), and selects the code number table ( 8) one of the code number table (8a
) to select the selector switch (7).

And, the changeover switch (11) has a code number table (8a
), the vector codeword in the codebook (10) pointed to by the code numbers sequentially read out by the changeover switch (9) is selected and outputted to the distance calculation means (12), and the distance Mjl is
'The calculation means (12) calculates the distance between the input vector (100) and the vector word 71 inputted via the changeover switch (11), and determines the optimum vector. It is output to the codeword selection means (13).

Further, the optimal vector first word selection means (13) selects the hector code that gives the minimum distance among the input distances, and outputs its n-th number through the output terminal (14). .

Note that from the perspective of roughly classifying input vectors using M intermediate vectors, the intermediate vectors do not have to be the same L-dimensional vectors as the vector codewords; By performing dimension reduction by selecting J dimensions in descending order of influence on the distance scale, the dimensionality of intermediate vectors can be reduced by adding almost no search error for the optimal vector codeword compared to conventional devices. It is possible to reduce the number by 7iJ.

Historically, the amount of memory can be reduced by reducing the number of dimensions of the intermediate codebook.

Furthermore, in the above embodiment, there is only one intermediate code length (3
), but regardless of this, even for high-speed exploration devices with a multi-stage tree structure of t, 1, distance calculations other than the final stage should be similarly reduced to the next-largest distance calculation. It is possible to speed up the search process.

Furthermore, a high-speed search can be realized by dividing the codebook itself and making each intermediate code correspond to each other, without having a code number table corresponding to each intermediate vector.

[Effects of the Invention] As explained above, according to the present invention, a multiple (L) dimensional input vector is converted into a multiple (J) dimensional degenerate dimensional input vector with a reduced number of dimensions, and Since the configuration is configured to calculate the distance to each intermediate vector question in the intermediate codebook, the number of calculations can be reduced, and the search speed can be increased by NXL/(MXJ+KXL) compared to the full search method. Compared to the conventional optimal hector codeword search device,
Further speed-up can be achieved without adding search errors for the optimal vector codeword, and even in the case of M- and J-L/2, the concatenation rate is 4/3 times that of the conventional device. There is an effect that can be achieved.

[Brief explanation of drawings]

Is Fig. 1 the optimal vector for one embodiment of the present invention? FIG. 2 is a block diagram showing the configuration of a conventional optimal vector codeword searching device. In the figure, (1) and (la) are optimal vector codes. word search device, (3) intermediate codebook, (6) optimal intermediate vector selection means, (8) code number table, (12) distance calculation means, (13) optimal code word selection means, (15) (16) is a dimension reduction means, and (16) is a reduction second-order distance calculation means. In the figure, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Claims] An optimal vector code whose distance on a predetermined distance evaluation scale is the smallest from a codebook consisting of a finite number (N) of multiple (L)-dimensional vector codewords with respect to multiple (L)-dimensional input vectors. In an optimal vector codeword searching device for searching for words, dimension reduction means converts a plurality (L)-dimensional input vector into a plurality (J)-dimensional reduced-dimensional input vector with a reduced number of dimensions; ) intermediate codebook consisting of a plurality of (M) plural (J)-dimensional intermediate vectors smaller than ), and a degenerate dimensional distance calculation means for calculating the distance between the degenerate dimensional input vector and each intermediate vector within the intermediate code width. , optimal intermediate vector selection means for selecting an intermediate vector that gives the minimum distance among the distances calculated by the degenerate dimensional distance calculation means, A plurality of (M) codeword number tables in which code addresses of a plurality of (K) codeword sets that are close to each selected intermediate vector and smaller than the finite plurality (N) are listed, and an optimal intermediate vector. distance calculation means for calculating the distance between each vector code word specified by the code word number table corresponding to the intermediate vector selected by the vector selection means and the input vector; an optimal vector code word selection means for selecting and outputting a vector code word giving the minimum distance; , the distance between the degenerate dimensional input vector and each intermediate vector within the intermediate code width is calculated using an intermediate codebook consisting of a plurality (M) of plural (J)-dimensional intermediate vectors, which is less than a finite plurality (N). An optimal vector codeword search device characterized by being configured as follows.