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
  • 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
  • G10L19/032—Quantisation or dequantisation of spectral components
  • G10L19/038—Vector quantisation, e.g. TwinVQ audio
• • H—ELECTRICITY
  • H03—ELECTRONIC CIRCUITRY
  • H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
  • H03M7/00—Conversion of a code where information is represented by a given sequence or number of digits to a code where the same, similar or subset of information is represented by a different sequence or number of digits
  • H03M7/30—Compression; Expansion; Suppression of unnecessary data, e.g. redundancy reduction
  • H03M7/3082—Vector coding
• • 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
  • G10L19/04—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 predictive techniques
  • G10L19/06—Determination or coding of the spectral characteristics, e.g. of the short-term prediction coefficients
  • G10L19/07—Line spectrum pair [LSP] vocoders
• • 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
  • G10L2019/0001—Codebooks
  • G10L2019/0004—Design or structure of the codebook
• • 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
  • G10L2019/0001—Codebooks
  • G10L2019/0016—Codebook for LPC parameters

Definitions

• Line spectrum pair parameters are used to represent coarse spectral information as generated by linear predictive coding analysis.
• linear predictive coding analysis to line spectrum pair parameter transformation generates a well behaved rank- ordered set that typically spans in the range of 0.0 to 0.5.
• the correlation properties of line spectrum pairs make them attractive for various scalar and vector quantization techniques, which are used in several recent speech compression standards. For example, IS- 127 standard entitled "Enhanced Variable Rate Codec,
• This method reduces the amount of codebook memory by splitting the vector into codebook segments, as compared to a brute force 22 bit 10 element code vector design which would require 42 mega words of storage and weighted squared error calculations, however, the weighted split vector quantization scheme does not afford the same performance advantage as the brute force method.
• One of the reasons is that overlap occurs between vector segments resulting in some combinations of codebook data that do not preserve the LSP ordering property and are therefore not valid for given linear predictive coding coefficients. The effect is that code vectors in the codebook are effectively pruned thereby reducing vector quantization coding gain.
• FIG. 1 shows a sample distribution plot of LSP parameters. Since LSP's are given to be a strictly ascending, rank ordered set, the first element in the second code vector must be greater than the last element in the first code vector. If an optimum code vector for the first segment of a 3-3-4 WSVQ was found to be ⁇ , ffi ⁇ , 6% ⁇ : ⁇ 0.10,0.15,0.20 ⁇ , then the first element of the second code vector, ⁇ , must be greater than 0.20 ( ⁇ ; 4 >0.20). But due to statistics in the LSP training database and the independent training of each segment's codebook, there may be a large number of vectors (entries) in the codebook that do not meet this criteria. The effect is that code vectors in the codebook that do not meet this criteria are invalid and rejected as possible candidates by the encoder and the codebook is effectively pruned.
• a known method that attempts to address this problem involves a weighted split vector quantization scheme in which the codebook values are based on the difference between a current and previously quantized line spectrum pair.
• This method known as delta coding, generally eliminates the overlap of codebook elements (as long as the delta omega is strictly positive), however a similar pruning effect can occur when the respective delta values are relatively large. For example, if the last quantized LSP value u ⁇ ,_ x is 0.3, and the values for the current LSP delta codebook ⁇ , span (0.05, 0.30), then any codebook value above 0.2 would be invalid since it would result in an LSP value greater than 0.5.
• LPC coefficient quantization in speech coding is disclosed in IEEE Conference Proceedings from the 1995 International Conference on Acoustics, Speech and Signal Processing in an article entitled "Fast and Low-Complexity LSF Quantization using Algebraic Vector Quantizer” authored by Minjie Xie et al. (1995).
• 10 LSP's per frame are quantized using a four segment split vector quantizer with a maximum segment length of three.
• the components of only one segment represent absolute LSP values and the remaining segments represent LSP differences normalized in such a manner that for each segment the sum of the components add to a value which is not greater than one.
• a single stored codebook is used to provide codebook entries for the one segment representing absolute LSP values, while codebook entries for the remaining segments are determined in real-time using the codevector indices and an implicit lattice structure.
• lattice structures typically requires less memory and search complexity than systems using only stored codebooks.
• normalized LSP differences in such a system can eliminate the occurrence of pruning. However, it has been found that although normalization can help insure that all codebook entries are valid, normalization can realign data so that a change in properties from the original parameters can occur.
• normalized delta quantization values affects the quantization error by changing the distribution and other statistics of the normalized delta quantization values so that they are different from those of the absolute values from which they were derived.
• codebooks containing normalized delta entries the normalized delta LSP values are mapped to a fixed range. Since the values of the lower and upper normalization bounds for a given LSP to be normalized are a function of previously quantized frame data, the mapping from the absolute LSP value to the normalized range is non-linear. This nonlinear mapping (warping) has the affect of changing the distribution and other statistics of the normalized delta value so that they are different from the statistics of the absolute values from which they were derived. The differences can affect the design of the codebook entries and the resulting quantization error.
• This warping affect can result in either a gain or loss in performance depending upon the statistics of the data being considered.
• the gain due to the elimination of pruning should be greater than the loss due to warping in order for the use of normalized delta values to provide an overall reduction in quantization error. If loss due to warping negates the gain due to the elimination of pruning, then the use of absolute values may be preferred. Hence determining LSP parameters from normalized vectors can result in lower coding performance.
• the use of stored codebooks allows the codebook entries to be optimally fit to a set of training data having a statistical distribution which matches that of the data to be quantized, which serves to reduce the average quantization error.
• the performance of a trained, stored codebook will generally be better than the performance of a codebook based on a highly complex lattice structure system that has severe memory requirements.
• FIG. 1 is a graph illustrating line spectrum pair parameter distribution for a number of analyzed speech frames.
• FIG. 2 is a block diagram generally depicting a speech decoder incorporating one embodiment of the invention in a wireless communication system.
• FIG. 3 is an illustration generally depicting one example of a codebook structure incorporating normalized quantization in accordance with one embodiment of the invention.
• FIG. 6 is a block diagram generally depicting a speech decoder incorporating another embodiment of the invention in a wireless communication system wherein codebook memory contains a combination of both absolute LSP values and normalized delta quantization data.
• a method and system for providing split vector quantization for use in determining constrained ordered set values, such as line spectrum pair parameters to determine spectral parameters in a data compression system utilizes multiple codebooks containing delta coded constrained ordered set values, such as line spectrum pair values, that are normalized to an upper and lower bound such as an available dynamic range of a constraint space.
• the values that are quantized in the segmented codebooks represent a percentage of a distance between a last quantized constrained ordered set value, such as a line spectrum pair value, and another higher (or highest possible) constrained ordered set value, such as a line spectrum pair value. Therefore all values stored in codebook memory result in valid LSP values which eliminates codebook pruning.
• split vector quantization includes scalar quantization where the vector length is one.
• An LSP reconstructor accesses the normalized delta quantization data in a segmented codebook which are used to determine constrained ordered set values such as line spectrum pair parameters to decode speech or other data.
• the LSP reconstructor reconstructs received spectral parameters based on reconstruction code data, namely the normalized delta quantization data of line spectrum pair parameters obtained from the split vector reconstruction codebooks.
• the LSP reconstructor dynamically generates line spectrum pair parameters based on the normalized delta quantization data.
• the reconstructed parameters are then passed to a short term synthesis filter to decode data, such as speech.
• the disclosed system and method instead of storing the absolute values of the line spectrum pair parameters in segmented codebooks, stores and utilizes a combination of at least two absolute value vectors and at least one vector of normalized delta quantization data for use in spectral quantization.
• FIG. 2 illustrates a speech decoder 10 in a communication system, such as a radio telephone communication system which receives encoded input data 12 representing encoded speech through speech parameters 14 such as spectral parameters, to reconstruct speech that has been transmitted over an air interface or channel.
• the speech parameters 14 are decoded as known in the art by a parameter decoder.
• the speech decoder 10 includes for example a fixed codebook with associated vectors 16 as known in the art, a long term prediction block 18, as known in the art, and a short term prediction LPC synthesis filter 20.
• the short term prediction LPC synthesis filter 20 uses LPC information that was generated based on LSP data derived from using split vector LSP reconstruction codebooks 22a-22c.
• the split vector LSP reconstruction codebooks 22a-22c serve as a split vector reconstruction code source containing reconstruction code data representing normalized delta quantization data for determining line spectrum pair parameters or other constrained ordered set parameters.
• Each entry of reconstruction code data in the codebooks 22a through 22c correspond to valid data for determining constrained ordered set data, such as line spectrum pair parameters, to facilitate reconstruction of the received encoded input data 12 or other spectral parameter.
• the speech parameters 14 may include among other parameters, a fixed codebook index parameter 24, a fixed codebook gain parameter 26, a long term delay parameter 28, a long term prediction gain parameter 30, and a normalization delta quantization (NDQ) codebook index parameter 32.
• NDQ normalization delta quantization
• the NDQ codebook index parameter 32 indexes a vector in one or more of the split vector LSP reconstruction codebooks 22a-22c.
• Line spectrum pair parameters are reconstructed by an LSP reconstructor 34 based on the indexed normalized delta quantization data in the split vector LSP reconstruction codebooks 22a-22c.
• An LSP to LPC transformer 36 transforms the LSP parameters to LPC information as known in the art. The LPC information is then used by the conventional short term prediction LPC synthesis filter 20 along with long term prediction information to generate output speech.
• the fixed codebook index parameter 24 indexes an address to a set of stochastic code vectors in the fixed codebook 16 to determine the stochastic component of the encoded speech waveform.
• the fixed codebook gain parameter 26 represents how strong the energy is from the encoded speech.
• the fixed codebook gain parameter 26 is multiplied in multiplier (mixer) 27 with the output vector from the fixed codebook 16 to provide gain to the fixed codebook data.
• the long term prediction block 18 as known in the art receives the delay parameter 28 which is used to determine the pitch period of the speech wave form.
• the long term prediction block 18 also receives the gain parameter 30 to determine the amount of pitch as known in the art.
• the short term prediction LPC synthesis filter 20 may be a code excited linear predictor (CELP) which determines the spectral envelope of the encoded speech as known in the art using short term parameters obtained from linear predictive coding information.
• CELP code excited linear predictor
• the linear predictive coding filter coefficients are converted from and represented by the constrained ordered set of line spectrum pair parameters based on the normalized delta quantization data in the NDQ LSP reconstruction codebooks 22a-22c.
• the normalized delta quantization data in the segmented LSP reconstruction codebooks 22a-22c is used to determine (reconstruct) the absolute value of line spectrum pair parameter values necessary for the short term prediction LPC synthesis filter 20.
• suitable normalized delta quantization data, A ⁇ , entries are determined for split vector LSP reconstruction codebooks 22a-22c such that each entry can be used for determining a valid LSP parameter, as shown in block 50. This may be done by a computer based on training speech frames or other suitable information that will be decoded.
• the delta coded LSP is normalized to a range, such as the available dynamic range, of the constraint space. That is, the codebook entries represent the percentage of the distance between a lower bound and an upper bound.
• the lower bound may be set to a previously obtained quantized value of a smaller LSP from the same speech frame, or to the minimum possible value of 0.0.
• the upper bound may be set to a previously obtained quantized value of a larger LSP from the same speech frame, or to the maximum possible value of 0.5. This can be expressed as:
• the normalized delta quantization data is stored in the split vector codebooks 22a-22c of the decoder 10. This may be done during decoder manufacture, on demand in the field or any suitable point in time.
• the decoder 10 receives speech parameters 14 for spectral information relating to the original speech as shown in block 54.
• the LSP reconstructor 34 based on the NDQ index parameters 32, retrieves a code vector from the split vector LSP reconstruction codebooks 22a-22c.
• the code vector from the split vector LSP reconstruction codebooks 22a-22c contains a representation of estimated LSP parameters. This is shown in block 56.
• the LSP reconstructor 34 reconstructs the LSP parameter based on the normalized delta quantization data from the codebooks 22a-22c as shown in block 58.
• the LSP ⁇ l is reconstructed from the codebook value ⁇ , and the previously obtained lower and upper bounds by:
• hybrid reconstruction codebooks 80a-80c are split vector codebooks. Codebooks 80a and 80c store at least two split vector segments (groups) containing absolute split vector code source data, such as absolute LSP values. Codebook 80b stores at least one segment of reconstruction code data representing normalized delta quantization of a set of ordered parameters, such as LSP parameters.
• the reconstruction code data representing normalized delta quantization of a set of ordered parameters is optimized based on a set of training data.
• the hybrid reconstruction codebook 80a-80c can afford advantages over known lattice codebook structures and those shown in FIG. 2 by employing at least two absolute vectors .
• An LSP reconstructor 82 reconstructs the set of ordered parameters based on both the at least two sets of absolute split vector code source data and the reconstruction code data representing normalized delta quantization of the set of ordered parameters.
• the LSP reconstructor uses the at least two segments of absolute split vector code source data and the at least one segment of normalized delta quantization data to convert the normalized quantization data to absolute quantization data for use by the LSP to LPC transformer 36.
• the reconstruction code data representing normalized delta quantization of the set of ordered parameters includes split vector codebook values quantized to represent a distance percentage between two absolute quantized line spectrum pair parameter values or other previously obtained upper and lower bounds.
• codebook 80a may contain codebook entries for quantizing one segment or group containing LSP's 1-3
• codebook 80c may contain LSP's 7-10
• codebook 80b may contain LSP's 4-6.
• codebooks 80a and 80c contain absolute LSP values while codebook 80b contains normalized delta quantization data.
• the lower bound Cut is set equal the quantized value of LSP 3
• the upper bound ⁇ is set to the quantized value of LSP 7. From FIG. 1 it can be observed that there is very little overlap in the distributions of LSP 3 and LSP 7. Thus, the loss due to pruning the codebook 80c is small.
• codebook 80c contains absolute LSP values.
• codebook 80b it can be seen from FIG. 1 that there is a significant overlap in the distributions of LSP 6 and LSP 7 and the distributions of LSP 3 and LSP 4. Therefore, there would be a significant loss due to pruning if absolute LSP values were used for codebook 80b.
• the use of normalized delta quantization data in codebook 80b results in a coding gain due to eliminating the pruning loss which is greater than a loss due to warping.
• fourteen LPC's can be used per frame to represent the coarse spectral information.
• This set of fourteen LPC's is converted to an ordered set of fourteen LSP's, ranging from 0.0 to 0.5.
• a four segment split vector quantizer having four codebooks is used.
• the first codebook contains entries representing the absolute values of LSP's one to three.
• the second codebook contain entries representing the absolute values of LSP's eleven to fourteen.
• the third codebook contains entries representing normalized delta LSP values derived from LSP's seven to ten. In computing these normalized delta quantization values, the lower bound is set equal to the quantized value of LSP 3 and the upper bound is set equal to the quantized value of LSP 11.
• the fourth codebook contains entries representing normalized delta LSP values derived from LSP's four to six.
• the lower bound is set equal to the quantized value of LSP 3
• the upper bound is set to the quantized value of LSP 7.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Spectroscopy & Molecular Physics (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)
• Theoretical Computer Science (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)

Applications Claiming Priority (3)

Publications (2)

Family

ID=21809582

Family Applications (1)

Country Status (4)

Families Citing this family (2)

Citations (1)

Family Cites Families (5)

• 1999
  • 1999-02-04 KR KR1020007008810A patent/KR20010040902A/ko not_active Application Discontinuation
  • 1999-02-04 EP EP99905742A patent/EP1062657A4/de not_active Withdrawn
  • 1999-02-04 WO PCT/US1999/002431 patent/WO1999041736A2/en not_active Application Discontinuation
  • 1999-02-04 JP JP2000531838A patent/JP2002503834A/ja active Pending

Patent Citations (1)

Non-Patent Citations (4)

Also Published As

Similar Documents

Legal Events