GB2300548A

GB2300548A - Vector quantization method for a communications system

Info

Publication number: GB2300548A
Application number: GB9508937A
Authority: GB
Inventors: Timor Kadir; Kwan Yee Lee; Mark Edward Adey
Original assignee: Motorola Ltd
Current assignee: Motorola Solutions UK Ltd
Priority date: 1995-05-02
Filing date: 1995-05-02
Publication date: 1996-11-06
Anticipated expiration: 2015-05-02
Also published as: GB9508937D0; GB2300548B

Abstract

A method for obtaining a reference signal in response to an input signal in a communications system wherein the reference signals (fixed vectors) are arranged in an original element list (vector quantization codebook) (62), comprises the steps of dividing the original element list (62) into at least two sub-lists (VQ sub-codebooks) (60, 64) and representing at least one reference signal (76) from the original element list (62) into the at least two sub-lists (60, 64).

Description

METHOD FOR A COMMUNICATIONS SYSTEM Field of the Invention This invention relates to communications systems and more particularly to vector quantization in a communications system.

Background of the Invention Many voice communications systems, such as the Trans-European Trucked Radio (TETRA) system for private mobile radio users, use speech processing units to encode and decode speech patterns. In such voice communications systems the speech encoder converts the analogue speech pattern into a suitable digital format for transmission and the speech decoder converts a received digital speech signal into an appropriate analog speech pattern.

As spectrum for such voice communications systems is a valuable resource, it is desirab to limit the channel bandwidth used, to maximise the number of users per frequency band. Hence, the primary objective in the use of speech coding techniques is to reduce the occupied capacity of the speech patterns as much as possible, by use of compression techniques, without losing fic elity.

Traditional speech coding techniques have tended to use scalar quantization (SQ), in processing the speech signal, due to its simplicity and good performance when the communication rate (i.e. bit rate) is sufficiently high. However, at low bit rates, SQ is not practical as there is often less than 1 bit/sample available for coding the speech signal.

Therefore, to satisfy both the reduction in occupied capacity and yet retain a high degree of fidelity, more efficient and sophisticated quantization methods are required.

A popular solution in speech coding is the application of vector quantization (VQ). A prime incentive in using VQ can be found in Shannon's rate distortion theory, as known to those skilled in the art, which states that better performance can always be achieved by coding vectors instead of scalars. The process of vector quantization is to represent an input vector as a member of a set of fixed vectors. This set of fixed vectors is known as the VQ codebook. The fixed vector in the VQ codebook which best represents the input vector is found by exhaustively searching all members of the VQ codebook and selecting the fixed vector which gives the minimum distance measure (or Euclidean distance) between it and the input vector. This procedure requires that every fixed vector in the VQ codebook is searched in order to find the best representation of the input vector.Consequently searching a full VQ codebook is computationally expensive. Although VQ has been shown to be very attractive and efficient in many areas of speech coding, it is not without its drawbacks.

It is well known in the art that the use of VQ codebooks often requires substantial processing and memory capabilities. For these reasons a careful trade-off must be made to get the best performance (complexity versus storage) from the VQ codebook. In order to overcome this drawback fast search algorithms with lower memory requirements have been proposed. These include tree-searched VQ, multi-stage VQ and lattice VQ. Although most of these algorithms suffer some performance degradation, they nevertheless present an attractive alternative to the computational complexity of a full-search VQ codebook.

A popular method of reducing the complexity of a VQ codebook search has been to divide the full VQ codebook, of size 'M', into a number of VQ sub-codebooks. This method effectively segments the vector space of the original VQ codebook into a number of fixed vector groups. A centroid is then calculated for each group of'N' fixed vectors and each centroid is then used to categorise input vectors as being best quantized by a fixed vector in a particular VQ sub-codebook. When the input vector is received a search algorithm performs a minimum distance calculation to identify which centroid the input vector is closest to. The VQ sub-codebook of 'N' fixed vectors corresponding to the chosen centroid is then searched to find the optimal fixed vector.This selection is performed by choosing the fixed vector from the chosen VQ subcodebook which gives the minimum distance between itself and the input vector. By searching through only the 'N' fixed vectors contained in the VQ sub-codebook, a fraction of the total number of fixed vectors are searched, as N < < M, providing a substantial computational saving.

One problem associated with the VQ sub-codebook approach is that it is not bit exact, i.e. an input vector can be calculated to be best represented by a fixed vector residing in an adjacent VQ sub-codebook, resulting in an incorrect choice of the fixed vector and sub-optimal quantization of the input vector.

Thus it is desirable to have a method for achieving bit exactness when using a VQ sub-codebook approach to represent input vectors.

Summarv of the Invention According to the present invention, there is a method provided to obtain a reference signal in response to an input signal in a communications system wherein the reference signals are arranged in an original element list. The method includes the steps of dividing the element list into at least two sub-lists and representing at least one reference signal from the original element list into the at least two sub-lists.

A preferred embodiment of the invention will now be described, by way of example only, with reference to the drawings.

Brief Description of the Drawings FIG. 1 shows a two-dimensional representation of a vector quantization (VQ) codebook.

FIG. 2 shows a two-dimensional representation of a VQ subcodebook.

FIG. 3 shows a two-dimensional representation of a VQ subcodebook where sub-optimal vector quantization occurs.

FIG. 4 shows a two-dimensional representation of a VQ subcodebook with some elements being represented in more than one segment in accordance with a preferred embodiment of the invention.

FIG. 5 shows a process of searching VQ sub-codebooks within an original VQ codebook, in accordance with a preferred embodiment of the invention.

FIG. 6 is a flow chart for dividing an element list (original VQ codebook) into sub-lists (VQ sub-codebooks).

FIG. 7 is a flow chart detailing a method of duplicating certain elements from the original VQ codebook into more than one VQ subcodebook to provide an optimum reference signal.

Detailed Description of the Drawings Referring first to FIG. 1, a simple two-dimensional representation of a vector quantization (VQ) codebook 10 is shown. The VQ codebook comprises a number of fixed vectors 12.

In operation the VQ codebook 10 would typically comprise a number of dimensions having significantly more fixed vectors represented than shown in FIG. 1. The fixed vector in the VQ codebook 10 that best represents the input vector is found by exhaustively searching all fixed vectors of the VQ codebook 10 and selecting the fixed vector which gives the minimum Euclidean distance measure between itself and the input vector.

Referring now to FIG. 2 the two-dimensional VQ codebook of FIG.

1 is shown, wherein the vector space of the VQ codebook 10 has been divided into a number of VQ sub-codebooks 22, 24, 26 and 28, as known in the prior art. By way of example only, the VQ codebook 10 has been divided into four VQ sub-codebooks 22, 24, 26 and 28 (segments) by applying the segment boundaries 14 and 16 to the original VQ codebook 10. Centroids 32, 34, 36 and 38 are then determined for each of the four segments 22, 24, 26 and 28 respectively.

In operation the centroids are used to categorise the input vectors as being best represented by a fixed vector in one of the VQ subcodebooks 22, 24, 26 or 28. When an input vector is received, the centroid closest to the input vector is calculated. The VQ sub-codebook according to the chosen centroid is then searched to find the most optimal fixed vector. The selection of the most optimal fixed vector is accomplished by choosing the fixed vector from the VQ sub-codebook that gives the minimum distance between itself and the input vector.

By searching through a VQ sub-codebook, a fraction of the total number of fixed vectors are searched, providing a substantial computational saving.

Referring now to FIG. 3, a problem associated with the prior art of searching VQ sub-codebooks is shown. An input vector 40 is received in the vector space such that the input vector 40 is nearest (minimum Euclidean distance) to the centroid 38 of the third segment and is therefore allocated to the third VQ sub-codebook 28. However, the input vector 40 resides close to a segment boundary 14 and in this case is allocated to a non-optimal segment, as shown. The input vector 40 is nearest to fixed vector 42, which resides in the adjacent, second VQ subcodebook 24. An incorrect choice of the optimal fixed vector is made, resulting in a non-bit-exact solution and sub-optimal quantization of the input vector 40.

Referring now to FIG. 4, an arrangement of the VQ subcodebook's fixed vectors is shown, according to a preferred embodiment of the invention. All fixed vectors that reside near the segment boundaries are duplicated in adjacent VQ sub-codebooks. For example, fixed vector 42 from the second VQ sub-codebook 24 and fixed vector 52 from the fourth VQ sub-codebook 26 are also included in the third VQ sub-codebook 28, characterised by the third centroid 38. The VQ subcodebook 28 associated with the third segment consequently contains these two additional fixed vectors, 42 and 52. The duplication process of fixed vectors that reside near segment boundaries is continued throughout all of the VQ sub-codebooks.A further example of the duplication process occurs in the fourth segment wherein, a fixed vector 50, residing in a first segment 22, is also included in the fourth segment and consequently fourth VQ sub-codebook 26.

Advantageously the duplication of some fixed vectors into adjacent segments (VQ sub-codebooks) removes the inaccuracies associated with the non-optimal VQ searches of existing VQ subcodebooks.

The duplication of some fixed vectors in such a VQ sub-codebook approach would typically require only a small amount of extra memory.

In cases where the available memory is limited a method for using the improved VQ sub-codebook approach by searching the original VQ codebook is now described.

FIG. 5 details the method of searching an original VQ codebook (element list) 62 for fixed vectors from a VQ sub-codebook, in accordance with a preferred embodiment of the invention. Sets of tables are used, e.g. Table 1, 60, represents a first VQ sub-codebook and Table 2, 64, represents a second VQ sub-codebook, to specify which fixed vectors within the original VQ codebook 62 need to be searched when using a VQ sub-codebook searching method. The original VQ codebook 62 is shown with a number of shaded areas. The shaded areas according to the shaded block 80 represent fixed vectors from the first VQ subcodebook and the shaded areas according to the shaded block 82 represent fixed vectors from the second VQ sub-codebook. The first location in Table 1, 60, represents the position of the first fixed vector 74 of the first VQ sub-codebook from the original VQ codebook 62.Each subsequent location in Table 1, 60, contains two digits. The first digit represents the step size to the next fixed vector of the first VQ subcodebook within the original VQ codebook 62. The second digit represents the number of repetitions to be made at this step size to retrieve subsequent first VQ sub-codebook fixed vectors from the original VQ codebook 62. In a similar manner Table 2, 64, details the location of the first fixed vector 72 in the second VQ sub-codebook, with digits in the subsequent locations representing the step size and repetition rate for subsequent fixed vectors in the second VQ subcodebook within the original VQ codebook 62. The shaded block 76 represents a fixed vector occurring near a segment boundary between the first and second VQ sub-codebooks. This fixed vector is therefore duplicated in both VQ sub-codebooks.

Although the preferred embodiment is described as a 2dimensional VQ codebook it is within the contemplation of the invention that the VQ codebook can have a much larger number of dimensions, a larger number of segment boundaries, each containing a finite number of fixed vectors. It is also within the contemplation of the invention that other table arrangements can be used to organise the searching fixed vectors of VQ sub-codebook within the orignal VQ codebook.

FIG. 6 shows a flow chart for dividing an element list (original VQ codebook) into at least two sub-lists (VQ sub-codebooks) by determining a centroid of the element list; dividing the element list into at least two sublists and estimating centroids for each sub-list. The elements from the element list are then clustered into the sub-lists and determining the elements which are to be represented in at least two sub-lists. The process for dividing the original VQ codebook begins, in step 100, with the centroid of the original VQ codebook CO being computed, as shown in step 102. A number (N) of small perturbation vectors P(k) are then added to the original VQ codebook CO to form 'N' sub-centroids C(k) where C(k) = CO + P(k), as in step 104.For each fixed vector in the original VQ codebook, the distance D(k) to each sub-centroid C(k) is calculated and the fixed vector is assigned to the VQ sub-codebook with minimum distance (MIN[D(k)]), as shown in step 106. The number of fixed vectors E(k) within each VQ sub-codebook are then computed, as in step 108. The number of fixed vectors E(k) in each VQ sub-codebook are then compared and if they are not approximately equal, in step 110, the sub-centroids C(k) are re-computed using the VQ sub-codebook fixed vectors, as in step 112. The 'N' small perturbation vectors P(k) are then modified to equalise E(k), as in step 114, and the process to obtain equal numbers of fixed vectors in each VQ sub-codebook continues, as shown in step 106.If the number of fixed vectors E(k) in each VQ sub-codebook are approximately equal, as in step 110, the sub-centroids C(k) are re-computed using the VQ sub-codebook fixed vectors, as in step 116 and the initial VQ sub-codebooks are designed, as shown in step 118. The process of dividing the original VQ codebook into a number of VQ sub-codebooks with approximately equal numbers of fixed vectors (elements) is therefore completed, as shown in step 120.

FIG. 7 is a flow chart detailing a method of duplicating certain elements from the original VQ codebook into more than one VQ subcodebook to provide an optimum reference signal, in accordance with a preferred embodiment of the invention. The "duplication" flow chart of FIG.

7 would be appended to the flow chart of the "initial sub-codebook design" detailed in FIG. 6. The process for determining which elements from the original element list are to be represented in at least two sub-lists commences in step 130 of FIG. 7. A database of test vectors is formed, as in step 132, and a fixed vector (element) is retrieved from this database, as shown in step 134. For each retrieved fixed vector, the distance D(k) to each sub-centroid C(k) is calculated, as in step 136, and the closest matching subcentroid Cx is selected (i.e. the sub-centroid with the minimum D(k)), as shown in step 138. The fixed vector (V) is then vector quantized using the VQ sub-codebook associated with that particular sub-centroid Cx to provide the fixed vector index Kx, as in step 140.Concurrently with steps 136, 138 and 140, the test vector is quantized using the original VQ codebook to provide an optimal fixed vector index KopX as shown in step 142. The two fixed vector indexes, Kop and Kx, are then compared, as in step 144. If Kop is equal to Kx then the fixed vector is compared to the last test vector from the database, as shown in step 148. If Kop is not equal to Kx then the original VQ codebook fixed vector provided by Kop is allocated to the VQ sub-codebook associated with Cx, as in step 146. The fixed vector is then compared to the last test vector from the database, as shown in step 148. If there are more fixed vectors available from the database, another fixed vector is retrieved, as shown in step 134, and the allocation process continues. If the fixed vector was the last from the test database, as in step 148, the duplication ratio is then calculated and checked to see whether it is acceptable or not, as in step 150. If the duplication ratio, as in step 150, is not acceptable the initial VQ sub-codebook, from the flowchart of FIG. 6 is redesigned, as shown in step 152. If the duplication ratio is acceptable, as in step 150, the design of the bit-exact VQ codebook is complete, as shown in step 154. The process for determining which elements from the original element list are to be represented in at least two sub-lists is completed in step 156.

The VQ codebook is, in the preferred embodiment, the TETRA algebraic code excited linear prediction (ACELP) codebook, as defined in document ETSI RES 06.20 from the European Telecommunications Standard Institute, F06921, Sophia Antipolis, France.

Thus, a method for achieving bit exactness when using a VQ subcodebook approach to represent input vectors is provided.

Claims

1. A method for providing a reference signal in response to an input signal in a communications system wherein reference signals are arranged in an original element list, the method comprising the steps of: dividing the original element list into at least two sub-lists; and representing at least one reference signal from the original element list into the at least two sub-lists.

2. The method of claim 1 wherein the step of dividing the original element list includes: determining a centroid of the original element list; dividing the original element list into at least two sub-lists; estimating centroids for each sub-list; clustering reference signals from the original element list into the at least two sub-lists; and determining which reference signals of the original element list are to be represented in at least two sub-lists.

3. The method of any of the preceding claims wherein the reference signals represented in the at least two sub-lists are the reference signals which occur near segment boundaries of the original element list.

4. The method of claim 1 wherein the original element list is a vector quantization (VQ) codebook and the reference signals within the original element list represent fixed vectors.

5. The method of claim 4 wherein the original VQ codebook is the TETRA ACELP speech codebook.

6. The method of any of the preceding claims wherein the sub-lists include approximately equal number of reference signals.

7. The method of any of the preceding claims wherein the sub-lists comprise a sub-list table and pointers to the original element list; the sublist table having step size information and information on the number of steps to be performed at this step size for pointing to sub-list reference signals within the original element list.

8. A method for achieving bit exactness when using a VQ sub-codebook approach to representing input vectors as substantially described with reference to FIG. 7.