EP1141944A2 - Method and arrangement for finding an optimal reconstruction point - Google Patents

Method and arrangement for finding an optimal reconstruction point

Info

Publication number
EP1141944A2
EP1141944A2 EP99964900A EP99964900A EP1141944A2 EP 1141944 A2 EP1141944 A2 EP 1141944A2 EP 99964900 A EP99964900 A EP 99964900A EP 99964900 A EP99964900 A EP 99964900A EP 1141944 A2 EP1141944 A2 EP 1141944A2
Authority
EP
European Patent Office
Prior art keywords
reconstruction
point
distance
points
reconstruction points
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP99964900A
Other languages
German (de)
French (fr)
Inventor
Morgan Lindqvist
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Telefonaktiebolaget LM Ericsson AB
Original Assignee
Telefonaktiebolaget LM Ericsson AB
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Telefonaktiebolaget LM Ericsson AB filed Critical Telefonaktiebolaget LM Ericsson AB
Publication of EP1141944A2 publication Critical patent/EP1141944A2/en
Withdrawn legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
    • G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders

Definitions

  • the invention relates generally to vector quantization and, more specifically, to a method and an arrangement for finding an optimal reconstruction point for an N- dimensional point to be quantized.
  • Vector quantization is computationally complex in the sense that a search for the optimal vector has to be carried out.
  • the optimal vector is usually defined as the vector giving the least residual error. Solutions with a "nearest neighbor" list or similar approaches to find the reconstruction points will reduce the complexity of the search.
  • One technical field where vector quantization is used is speech coding, where several correlated gain parameters are jointly quantized.
  • All CELP (Code Excited Linear Prediction) coders of today minimize the mean square error in the weighted speech domain when quantizing many of the parameters.
  • the correlated gain parameters should also be quantized with regard to the distortion in the weighted speech domain and not to the traditional distortion in the gain domain. This makes it impossible to create a "nearest neighbor" list since this list with this distortion measure, is signal dependent, i.e. depends on other values than the ones to be quantized.
  • the distortion measure is signal dependent, all reconstruction points have to be searched in order to find the best reconstruction point.
  • D-AMPS Digital American Mobile Phone System
  • the complexity of the search will be high since all reconstruction points have to be searched. It will be especially complex for the above mentioned CELP coders, since the distortion calculation itself is complex.
  • the object of the invention is to reduce the complexity of the current methods and apparatuses for finding reconstruction points in connection with vector quantization.
  • Fig. 1 illustrates an example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are strongly correlated
  • Fig. 1 illustrates a relationship between predicted values and sorted reconstruction points for the reconstruction points illustrated in Fig. 1,
  • FIG. 3 illustrates another example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are evenly distributed, and
  • - Fig. 4 illustrates a further example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are grouped along the border of a circle.
  • Fig. 1 illustrates an example of reconstruction points for transformed gain parameters resulting from a speech codec.
  • the transformation has been done to reduce the variance of the gain. This reduced variance makes it possible to construct a quantizer that introduces less distortion. As apparent from Fig. 1, the two different gains are strongly correlated.
  • the reconstruction points in Fig. 1 are sorted in advance by ordering them by their distance from a reference point in an N- dimensional space.
  • the reconstruction points for correlated data are grouped around a line or curve in the N-dimensional space, it is possible to create a cross-reference list that makes it possible to access a set of reconstruction points surrounding an arbitrarily predicted reconstruction point.
  • the reconstruction points are grouped around a line in the N-dimensional space as in Fig. 1, it is possible to create the cross- reference list by sorting the reconstruction points by their distance to the reference point at one end of the line.
  • a transfer function for transforming the prediction value, in this example the "distance to reference point", into “sorted reconstruction point” is shown.
  • the transfer function in Fig. 2 which is created in advance, is used to predict a reconstruction point, called “sorted VQ index” in Fig. 2, from the "distance to reference point” as will be described below.
  • the transfer function is created by calculating the distance to the reference point, as described below, for each sorted reconstruction point.
  • the transfer function in Fig. 2 can very well be represented by two or three linearized first degree equations, giving a low complexity transformation function from "distance to reference point” into “sorted reconstruction point” .
  • finding of an optimal reconstruction point for an N-dimensional point to be quantized, where the N-dimensional point corresponds to N parameters extracted from an input signal is done in the following way:
  • N is the dimension of the space
  • X t is the unquantized value in the i:th dimension
  • Refi is the reference point in the i:th dimension.
  • the size of the selected set of reconstruction points should be chosen large enough so that the optimal reconstruction point always exists within the selected set.
  • the size of the selected set of reconstruction points may in fact be predetermined.
  • the distortion value represents the difference between an original signal, e.g. a speech signal coming in to a speech encoder, and a reconstructed signal, e.g. an output speech signal from the speech encoder.
  • the search complexity will be reduced by approximately 80% since only a small part of the total number of reconstruction points has to be searched.
  • the complexity can be further reduced if it can be accepted that the optimal reconstruction point is not always used. This further reduction of complexity can be achieved by reducing the selected set of reconstruction points.
  • the only cost for the prediction is a small memory increase required for storing the cross reference list used to access the reconstruction points sorted according to the distance measure.
  • This method is called index assignment. If index assignment is unnecessary for the reconstruction points, the cross reference list is not needed since the reconstruction points can be ordered in this way from the beginning. Thus, no extra memory is required.
  • the reconstruction points are grouped almost evenly in a square in a two dimensional space.
  • the reference point is therefore placed in a smaller dimension where it is possible to find correlation. In this example, this is the first dimensional space.
  • the reference point may be placed at the x-mark to the right, and only the distance in the horizontal plane would be used. This will lead to that the selected set of reconstruction points will be larger than it would have been, had there been a correlation between the values in the different dimensions. However, the saving in complexity will still be considerable.
  • the reconstruction points are approximately placed along the border of a circle.
  • the "distance" can be the angular distance, i.e. the angle, to the predicted reconstruction point. If the predicted reconstruction point in this case has an angle close to 0 or 2 ⁇ radians, the selected set of reconstruction points must include reconstruction points that have an angle close to both 0 and 2 ⁇ . This will lead to a low complexity search also when the reconstruction points are placed in this manner.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

To reduce the complexity of finding reconstruction points in connection with vector quantization, a set of reconstruction points close to a predicted reconstruction point is selected, whereupon a distortion measure is calculated for all the reconstruction points of the selected set, and that reconstruction point that gives the smallest distortion, is chosen as the optimal reconstruction point. Since only a small number of reconstruction points has to be searched, the complexity of the search is significantly reduced.

Description

METHOD AND ARRANGEMENT FOR FINDING AN OPTIMAL RECONSTRUCTION POINT
TECHNICAL FIELD The invention relates generally to vector quantization and, more specifically, to a method and an arrangement for finding an optimal reconstruction point for an N- dimensional point to be quantized.
BACKGROUND OF THE INVENTION When data are to be quantized, vector quantization is often used. This is due to the fact that the distortion resulting from vector quantization always is less than the distortion resulting from scalar quantization. Moreover, the distortion will decrease further if the data to be quantized are correlated.
Vector quantization is computationally complex in the sense that a search for the optimal vector has to be carried out. The optimal vector is usually defined as the vector giving the least residual error. Solutions with a "nearest neighbor" list or similar approaches to find the reconstruction points will reduce the complexity of the search.
One technical field where vector quantization is used is speech coding, where several correlated gain parameters are jointly quantized. All CELP (Code Excited Linear Prediction) coders of today minimize the mean square error in the weighted speech domain when quantizing many of the parameters. To achieve the best speech quality, the correlated gain parameters should also be quantized with regard to the distortion in the weighted speech domain and not to the traditional distortion in the gain domain. This makes it impossible to create a "nearest neighbor" list since this list with this distortion measure, is signal dependent, i.e. depends on other values than the ones to be quantized. When the distortion measure is signal dependent, all reconstruction points have to be searched in order to find the best reconstruction point. One example of this is the gain quantization in the newly standardized speech codec IS-641 used in the Digital American Mobile Phone System (D-AMPS).
The complexity of the search will be high since all reconstruction points have to be searched. It will be especially complex for the above mentioned CELP coders, since the distortion calculation itself is complex.
SUMMARY OF THE INVENTION
The object of the invention is to reduce the complexity of the current methods and apparatuses for finding reconstruction points in connection with vector quantization.
This is attained by selecting a set of reconstruction points close to a predicted reconstruction point, evaluating a distortion measure for all reconstruction points of the selected set, and choosing, as reconstruction point, the reconstruction point of the selected set that gives the smallest distortion.
Since only a small number of the total number of reconstruction points has to be searched with this method, the complexity is significantly reduced.
BRIEF DESCRIPTION OF THE DRAWING
The invention will be described more in detail below with reference to the appended drawing, on which - Fig. 1 illustrates an example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are strongly correlated,
- Fig. 2 illustrates a relationship between predicted values and sorted reconstruction points for the reconstruction points illustrated in Fig. 1,
- Fig. 3 illustrates another example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are evenly distributed, and
- Fig. 4 illustrates a further example of reconstruction points for a 2-dimensional vector quantizer, where the values in the two dimensions are grouped along the border of a circle.
PREFERRED EMBODIMENTS
Fig. 1 illustrates an example of reconstruction points for transformed gain parameters resulting from a speech codec. The transformation has been done to reduce the variance of the gain. This reduced variance makes it possible to construct a quantizer that introduces less distortion. As apparent from Fig. 1, the two different gains are strongly correlated.
In accordance with the invention, the reconstruction points in Fig. 1 are sorted in advance by ordering them by their distance from a reference point in an N- dimensional space. In Fig. 1, that reference point is placed in the upper right-hand corner, and it is supposed that N = 2.
Since the reconstruction points for correlated data are grouped around a line or curve in the N-dimensional space, it is possible to create a cross-reference list that makes it possible to access a set of reconstruction points surrounding an arbitrarily predicted reconstruction point. In case the reconstruction points are grouped around a line in the N-dimensional space as in Fig. 1, it is possible to create the cross- reference list by sorting the reconstruction points by their distance to the reference point at one end of the line.
In Fig. 2, a transfer function for transforming the prediction value, in this example the "distance to reference point", into "sorted reconstruction point" is shown. The transfer function in Fig. 2 which is created in advance, is used to predict a reconstruction point, called "sorted VQ index" in Fig. 2, from the "distance to reference point" as will be described below. The transfer function is created by calculating the distance to the reference point, as described below, for each sorted reconstruction point.
The transfer function in Fig. 2 can very well be represented by two or three linearized first degree equations, giving a low complexity transformation function from "distance to reference point" into "sorted reconstruction point" .
In accordance with the invention, finding of an optimal reconstruction point for an N-dimensional point to be quantized, where the N-dimensional point corresponds to N parameters extracted from an input signal, is done in the following way:
Use the unquantized values of the N-dimensional point to be quantized to calculate the distance from the reference point by means of the equation
where N is the dimension of the space, Xt is the unquantized value in the i:th dimension, and Refi is the reference point in the i:th dimension.
Then, transform the calculated distance into "sorted reconstruction point" using the linearized version of the transfer function shown in Fig. 2.
After that, among the reconstruction points as sorted, select a set of reconstruction points surrounding the predicted reconstruction point.
The size of the selected set of reconstruction points should be chosen large enough so that the optimal reconstruction point always exists within the selected set.
This can be checked during development by conducting a search for the optimal reconstruction point both in a conventional manner, i.e. over all reconstruction points, and in accordance with the invention, and verifying that the same optimal reconstruction point always is found. In the example in Fig. 1, it will only be necessary to include about 10 % of the total number of reconstruction points on each side of the predicted reconstruction point to ensure that the optimal reconstruction point is included in the selected set of reconstruction points.
Thus, the size of the selected set of reconstruction points may in fact be predetermined.
Thereafter, for each reconstruction point within the selected set of reconstruction points, a so called distortion value is calculated.
The distortion value represents the difference between an original signal, e.g. a speech signal coming in to a speech encoder, and a reconstructed signal, e.g. an output speech signal from the speech encoder.
Should the distortion value be more dependent upon one dimension than the other dimension(s), this can be taken into account in the above equation for distance calculation, e.g. by giving different weight to the different dimensions.
Finally, as the optimal reconstruction point, choose the reconstruction point having the lowest calculated distortion value in said selected set of reconstruction points surrounding the predicted reconstruction point.
In the above example, the search complexity will be reduced by approximately 80% since only a small part of the total number of reconstruction points has to be searched.
The complexity can be further reduced if it can be accepted that the optimal reconstruction point is not always used. This further reduction of complexity can be achieved by reducing the selected set of reconstruction points. The only cost for the prediction is a small memory increase required for storing the cross reference list used to access the reconstruction points sorted according to the distance measure. When the reconstruction points are to be transmitted over a transmission channel with bit errors, the reconstruction points are often so ordered that a single bit error will give as small a distortion as possible. This method is called index assignment. If index assignment is unnecessary for the reconstruction points, the cross reference list is not needed since the reconstruction points can be ordered in this way from the beginning. Thus, no extra memory is required.
In the example in Fig. 3, the reconstruction points are grouped almost evenly in a square in a two dimensional space. In this example, there is no correlation between the values in the different dimensions. The reference point is therefore placed in a smaller dimension where it is possible to find correlation. In this example, this is the first dimensional space. In Fig. 3, the reference point may be placed at the x-mark to the right, and only the distance in the horizontal plane would be used. This will lead to that the selected set of reconstruction points will be larger than it would have been, had there been a correlation between the values in the different dimensions. However, the saving in complexity will still be considerable.
In Fig. 4, the reconstruction points are approximately placed along the border of a circle. In this case, the "distance" can be the angular distance, i.e. the angle, to the predicted reconstruction point. If the predicted reconstruction point in this case has an angle close to 0 or 2π radians, the selected set of reconstruction points must include reconstruction points that have an angle close to both 0 and 2π. This will lead to a low complexity search also when the reconstruction points are placed in this manner.
As should be apparent from the above, the complexity of finding the optimal reconstruction point will be reduced by means of the invention. It should also be obvious to anyone skilled in the art that the method described above may be run on a processor.

Claims

1. In an N-dimensional space, N>2, containing a predetermined number of predetermined reconstruction points to be used by a decoder to reconstruct an input signal coded by an encoder, a method of finding an optimal reconstruction point for an N-dimensional point to be quantized, said N-dimensional point corresponding to N parameters extracted from the input signal, characterized by
- sorting, in advance, the predetermined reconstruction points with regard to their distance from a reference point, - creating, in advance, a function to transform a distance from the reference point into a sorted reconstruction point,
- calculating the distance from the reference point to the N-dimensional point,
- transforming, by means of the created function, the calculated distance into a predicted reconstruction point, the distance of which from the reference point being closest to the distance to be transformed,
- selecting a set of reconstruction points surrounding, as sorted, the predicted reconstruction point,
- calculating a distortion value for all reconstruction points in said selected set of reconstruction points, and - choosing the reconstruction point having the lowest calculated distortion value as the optimal reconstruction point in said selected set of reconstruction points.
2. The method as claimed in claim 1, characterized in that a predetermined set of reconstruction points surrounding, as sorted, the predicted reconstruction point, is selected.
3. An arrangement for finding, in an N-dimensional space, N>2, containing a predetermined number of predetermined reconstruction points to be used by a decoder to reconstruct an input signal coded by an encoder, an optimal reconstruction point for an N-dimensional point to be quantized, said N-dimensional point corresponding to N parameters extracted from the input signal, characterized by
- means for sorting, in advance, the predetermined reconstruction points with regard to their distance from a reference point, - means for creating, in advance, a function to transform a distance from the reference point into a sorted reconstruction point,
- means for calculating the distance from the reference point to the N-dimensional point,
- means for fransforming, by means of the created function, the calculated distance into a predicted reconstruction point, the distance of which from the reference point being closest to the distance to be transformed,
- means for selecting a set of reconstruction points surrounding, as sorted, the predicted reconstruction point,
- means for calculating a distortion value for all reconstruction points in said selected set of reconstruction points, and
- means for choosing the reconstruction point having the lowest calculated distortion value as the optimal reconstruction point in said selected set of reconstruction points.
4. The arrangement as claimed in claim 3, characterized in that said means for selecting are adapted to select a predetermined set of reconstruction points surrounding, as sorted, the predicted reconstruction point.
EP99964900A 1998-12-21 1999-12-17 Method and arrangement for finding an optimal reconstruction point Withdrawn EP1141944A2 (en)

Applications Claiming Priority (3)

Application Number Priority Date Filing Date Title
SE9804455A SE9804455L (en) 1998-12-21 1998-12-21 Procedure and arrangement for finding an optimal reconstruction point
SE9804455 1998-12-21
PCT/SE1999/002404 WO2000038176A2 (en) 1998-12-21 1999-12-17 Method and arrangement for finding an optimal reconstruction point

Publications (1)

Publication Number Publication Date
EP1141944A2 true EP1141944A2 (en) 2001-10-10

Family

ID=20413781

Family Applications (1)

Application Number Title Priority Date Filing Date
EP99964900A Withdrawn EP1141944A2 (en) 1998-12-21 1999-12-17 Method and arrangement for finding an optimal reconstruction point

Country Status (6)

Country Link
EP (1) EP1141944A2 (en)
JP (1) JP2002533962A (en)
AU (1) AU3093000A (en)
CA (1) CA2356152A1 (en)
SE (1) SE9804455L (en)
WO (1) WO2000038176A2 (en)

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4670851A (en) * 1984-01-09 1987-06-02 Mitsubishi Denki Kabushiki Kaisha Vector quantizer
US5077798A (en) * 1988-09-28 1991-12-31 Hitachi, Ltd. Method and system for voice coding based on vector quantization
JP3151874B2 (en) * 1991-02-26 2001-04-03 日本電気株式会社 Voice parameter coding method and apparatus

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See references of WO0038176A3 *

Also Published As

Publication number Publication date
AU3093000A (en) 2000-07-12
CA2356152A1 (en) 2000-06-29
WO2000038176A2 (en) 2000-06-29
WO2000038176A3 (en) 2000-10-19
JP2002533962A (en) 2002-10-08
SE9804455L (en) 2000-06-22
SE9804455D0 (en) 1998-12-21

Similar Documents

Publication Publication Date Title
US6885988B2 (en) Bit error concealment methods for speech coding
US7325023B2 (en) Method of making a window type decision based on MDCT data in audio encoding
EP2915166B1 (en) A method and apparatus for resilient vector quantization
EP2441049A2 (en) Method for encoding/decoding a 3d mesh model that comprises one or more components
CN109313908A (en) Audio coder for being encoded to audio signal, the method for being encoded to audio signal and the computer program for considering the spike spectral regions detected in upper frequency band
JP2004526213A (en) Method and system for line spectral frequency vector quantization in speech codecs
US5487087A (en) Signal quantizer with reduced output fluctuation
US7283968B2 (en) Method for grouping short windows in audio encoding
US5748839A (en) Quantization of input vectors and without rearrangement of vector elements of a candidate vector
US7739119B2 (en) Technique for implementing Huffman decoding
Slyz et al. A nonlinear VQ-based predictive lossless image coder
EP1141944A2 (en) Method and arrangement for finding an optimal reconstruction point
KR950003557B1 (en) Encoding method of voice sample and signal sample
US20110112841A1 (en) Apparatus
Ho et al. A variable-rate image coding scheme with vector quantization and clustering interpolation
RU2461079C2 (en) Simplified vector indexing and deindexing
JPH03171828A (en) Compression encoder
KR20180026528A (en) A bit error detector for an audio signal decoder
JP2890522B2 (en) Method and apparatus for adaptive transform coding
Bruhn et al. The FlexCode speech and audio coding approach
JP2638209B2 (en) Method and apparatus for adaptive transform coding
JP2923996B2 (en) Method and apparatus for adaptive transform coding
Torres et al. Improvements on stochastic vector quantization of images
Merazka et al. Robust split vector quantization of LSP parameters at low bit rates
Fonollosa et al. Adaptive multistage vector quantization

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20010613

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE

AX Request for extension of the european patent

Free format text: AL;LT;LV;MK;RO;SI

RIN1 Information on inventor provided before grant (corrected)

Inventor name: LINDQVIST, MORGAN

RAP1 Party data changed (applicant data changed or rights of an application transferred)

Owner name: TELEFONAKTIEBOLAGET LM ERICSSON (PUBL)

RBV Designated contracting states (corrected)

Designated state(s): DE GB

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20040701