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
  • G10L19/04—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
  • G10L2019/0005—Multi-stage vector quantisation
• • 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
  • G10L2019/0007—Codebook element generation

Definitions

• This invention relates to speech coding and methods of optimising the performance of speech codecs in communications systems.
• the invention is applicable to, but not limited to, speech codecs that accommodate wideband and narrowband speech signals without compromising the overall performance of the speech codec quantiser.
• GSM global system for mobile communications
• TETRA TErrestrial Trunked RAdio
• a 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 fidelity.
• VQ vector quantisation
• the process of vector quantisation 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.
• VQ has been shown to be very attractive and efficient in many areas of speech coding, it is not without its drawbacks .
• Wideband speech codecs are likely to find application in telephone conferencing.
• Wideband speech codecs have an input speech bandwidth covering the 50Hz to 7KHz range, compared to narrowband or telephone-band codecs that have an input speech bandwidth of 250Hz to 3.3KHz.
• Tandemming is a term which is used to describe the situation where speech previously processed by one speech encoder/decoder is processed by a second speech encoder/decoder pair.
• the speech quality requirement of such tandemmed codecs is to achieve equivalence to the best narrowband codecs i.e. GSM Enhanced full-rate codec (EFR) . It is therefore appropriate to consider the performance of any wideband line spectral frequency (LSF) VQ scheme in the presence of narrowband speech.
• GSM Global System for Mobile Communications
• one option may be to use a classified VQ scheme with two sets of codebooks: one to represent the wideband speech and one to represent the narrowband speech.
• a respective codebook would be selected by a special "mode" bit, where the mode bit indicates whether the subsequent data bits represent a wideband or narrowband speech signal.
• the representative codecs have been simulated with each predictor of the speech codec arranged to be an 18 th order split vector quantiser, with the eighteen associated line spectral frequencies split into six groups of three bits each.
• Table 1 7KHz & 3KHz Spectral Distortion Results for the 1st order MA-PVQ 40 bit Quantisers.
• Table 1 details the wideband and narrowband spectral distortion figures for a 40-bit first order moving average quantiser trained on 50:50 wideband: narrowband speech.
• the configuration column denotes the number of bits allocated to each of six, moving- average predictive split-vector quantisers, applied to LSFs 1-3, 4-6, 7-9, 10-12, 13-15 and 16-18 respectively.
• a wideband speech codec would typically be represented by an even distribution of bits allocated to each of the six split-vector quantisers, to provide an approximately even frequency response across the full range of the line spectral frequencies.
• a narrowband speech codec would have an uneven distribution of bits associated with each quantiser, with more bits allocated to the lower frequencies of the LSFs.
• the present invention aims to provide a speech codec and method of optimising a performance of the speech codec to at least alleviate some of the aforementioned disadvantages .
• a speech coder for a speech communications unit in accordance with claim 1 is provided.
• a speech communications unit adapted to include the speech coder of any one of claims 1 to 10 is provided.
• a method of generating a speech vector codebook in a speech communications unit in accordance with claim 12 is provided.
• a method of encoding a speech signal in accordance with claim 22, is provided.
• a speech communications unit adapted to employ a speech encoding method in accordance with any one of claims 22 to 26 is provided.
• a method of decoding a speech signal in accordance with claim 28, is provided.
• a speech communications unit adapted to employ a speech decoding method in accordance with any one of claims 28 to 31 is provided.
• FIG. 1 shows a block diagram of a code excited linear predictive speech encoder that can be adapted to support the various inventive concepts of a preferred embodiment of the present invention
• FIG. 2 shows a block diagram of a code excited linear predictive speech decoder that can be adapted to support the various inventive concepts of a preferred embodiment of the present invention
• FIG. 3 shows a 2-way split VQ codebook applied to eighteen wideband line spectral frequencies (LSFs) adapted to support the inventive concepts of the preferred embodiments of the present invention
• FIG. 4 shows a preferred packing arrangement of bits for the 2-way split VQ codebook of FIG. 3, adapted to support the inventive concepts of the preferred embodiments of the present invention
• FIG. 5 shows an octagon partitioned to reflect eight separate locations for the 2-way split VQ codebook of FIG. 3, in accordance with a preferred embodiment of the present invention
• FIG. 6 shows a graph that demonstrates the error resilience of a preferred embodiment of the present invention in the presence of 10% bit errors applied on a per-bit basis with natural ordering
• FIG. 7 shows a further graph that demonstrates the error resilience of a preferred embodiment of the present invention in the presence of 10% bit errors applied on a per-bit basis with ranked ordering.
• FIG. 1 a block diagram of a code excited linear predictive speech encoder 100, according to a preferred embodiment of the present invention, is shown.
• An acoustic input signal to be analysed is applied to speech coder 100 at microphone 102.
• the input signal is then applied to filter 104.
• Filter 104 will generally exhibit band-pass filter characteristics. However, if the speech bandwidth is already adequate, filter 104 may comprise a direct wire connection.
• the analog speech signal from filter 104 is then converted into a sequence of N pulse samples, and the amplitude of each pulse sample is then represented by a digital code in analog-to-digital (A/D) converter 108, as known in the art.
• the sampling rate is determined by sample clock (SC) .
• the sample clock SC is generated along with the frame clock (FC) via clock 112.
• A/D 108 which may be represented as input speech vector s (n)
• coefficient analyser 110 The digital output of A/D 108, which may be represented as input speech vector s (n) , is then applied to coefficient analyser 110.
• This input speech vector s (n) is repetitively obtained in separate frames, i.e., blocks of time, the length of which is determined by the frame clock (FC) , as is known in the art.
• LPC linear predictive coding
• the generated speech coder parameters may include the following: LPC parameters, long-term predictor (LTP) parameters, excitation gain factor (V) (along with the best excitation codeword I) .
• LPC parameters are applied to multiplexer 150 and sent over the channel 152 for use by the speech synthesizer at the decoder.
• the input speech vector s (n) is also applied to subtractor 130, the function of which is described later.
• coefficient analyser 110 has been adapted to incorporate the specially constructed family of embedded codebooks.
• the codebook search controller 140 selects the best indices and gains from the adaptive codebook within block 116 and the stochastic codebook within block 114 in order to produce a minimum weighted error in the summed chosen excitation vector used to represent the input speech sample.
• the output of the stochastic codebook 114 and the adaptive codebook 116 are input into respective gain functions 122 and 118.
• the gain-adjusted outputs are then summed in summer 120 and input into the LPC filter 124, as is known in the art. For each individual excitation vector u ⁇ (n), a reconstructed speech vector s'i(n) is generated for comparison to the input speech vector s(n).
• Gain block 122 scales the excitation gain factor W .
• Such gain may be pre-computed by coefficient analyser 110 and used to analyse all excitation vectors, or may be optimised jointly with the search for the best excitation codeword I, generated by codebook search controller 140.
• the scaled excitation signal ui (n) is then filtered by the linear predictive coding filter 124, which preferably includes a long-term predictor (LTP) filter and a short- term predictor (STP) filter, to generate the reconstructed speech vector s'i(n).
• the reconstructed speech vector s'i(n) for the i-th excitation code vector is compared to the same block of input speech vector s (n) by subtracting these two signals in subtractor 130.
• the difference vector ⁇ i (n) represents the difference between the original and the reconstructed blocks of speech.
• the difference vector is perceptually-weighted by weighting filter 132, utilising the weighting filter parameters (WTP) generated by coefficient analyser 110. Perceptual weighting accentuates those frequencies where the error is perceptually more important to the human ear, and attenuates other frequencies.
• An energy calculator function inside the codebook search controller 140 computes the energy of the weighted difference vector e'i(n).
• the codebook search controller compares the i-th error signal for the present excitation vector Ui(n) against previous error signals to determine the excitation vector producing the minimum error.
• the code of the i-th excitation vector having a minimum error is then output over the channel as the best excitation code I .
• codebook search controller 140 may determine a particular codeword that provides an error signal having some predetermined criteria, such as meeting a predefined error threshold.
• the coefficient analyzer 110 has also been adapted to employ at least some of the inventive concepts of the present invention. To accommodate vectors in either the wideband or narrowband vector space, the coefficient analyser 110 is used to train the quantisers and to determine whether the input speech comprises wideband or narrowband speech.
• the inventors of the present invention have recognised the opportunity to use the same training data, or at least very similar data, to train each of the quantisers.
• the different sized quantisers cover much the same signal vector space and hence a smaller quantiser is embedded within the larger quantiser leading to a more compact representation.
• the quantiser set will preferably refer to a wideband or narrowband arrangement.
• the codebook index transmission is structured in order to minimise the effect of errors to this mode bit, as described later with respect to FIG. 4.
• the consequence of such a careful structuring of the codebook index transmission means that any bit error (s) in the mode bit have much less impact than in any two or more independent-codebook prior art approach.
• FIG. 2 a block diagram of a code excited linear predictive speech decoder 200 is shown, according to a preferred embodiment of the present invention.
• the decoder functionality is substantially the reverse of that of the encoder.
• the received multiplexed signal is input into demultiplexer 202, which separates the excitation parameters 204 from the LPC parameters 206.
• LPC linear predictive coding
• the LPC parameters are input into an LPC de-quantiser, stability check and correction block 210 to obtain a local stable version of the synthesis filter even in the presence of channel bit errors.
• the LPC de-quantise, stability check and correction block 210 has been adapted to encompass the inventive concepts contained herein.
• the LPC de-quantise, stability check and correction block 210 receives the LPC parameters and mode bit sent from the corresponding encoder function.
• the LPC de-quantise function of block 210 includes the corresponding embedded codebook arrangement of the encoder, such that the determination of the at least one mode bit can select the embedded codebook arrangement that best describes the encoded and transmitted speech signal.
• the LPC de- quantise, stability check and correction block 210 also controls the filter co-efficients of the LPC synthesis filter 222 in order to reconstruct the transmitted speech vector s'i(n).
• the output from the LPC synthesis filter 222 is input to a post filter process 224, which subsequently outputs the reconstructed speech 226.
• the excitation parameters 204 may include: excitation gain factor together with the best excitation codeword I, and are input into an adaptive non-linear smoothing function 208.
• the output from the adaptive non-linear smoothing function 208 provides the precise adaptive and stochastic codebook indices and gains that form the excitation for the synthesis filter. As such, the outputs from the adaptive non-linear smoothing function 208 are input to stochastic codebook 218 and adaptive codebook 212.
• the gain controls are input to adaptive codebook gain block 214, which receives an output from the adaptive codebook 212, and stochastic codebook gain block 220, which receives an output from the stochastic codebook 218.
• the output from the respective gain blocks 214, 220 are input to summing junction 216, whose output is fed into the LPC synthesis filter 222 and fed back to the adaptive codebook 212, as known in the art.
• FIG. 3 shows a 2-way split VQ codebook applied to eighteen wideband line spectral frequencies (LSFs) .
• LSFs wideband line spectral frequencies
• the input LSFs (L1-L18) 250 are quantised by first quantiser 254 and second quantiser 268 to derive estimates and the two binary indices "II" 270 and "12" 272 using a respective first embedded codebook 256 and second embedded codebook 262.
• the LSFs (L1-L18) 250 are fed into a mode-bit detector 252, that selects the respective embedded codebook to provide the most appropriate one for the speech signal presented.
• the first embedded codebook 256 contains a first set of core entries, in this case appropriate for wideband speech 260, and additional entries appropriate for narrowband speech 258.
• the second embedded codebook 262 contains a second set of core entries, this time appropriate for narrowband speech 266, and additional wideband entries 264.
• the core entries are always searched in each quantiser and are indexed by a set of core bits. Additional entries are searched, depending upon the mode, and a set of "extra" bits are formed.
• the codebook is structured such that when the full codebook is searched, the "extra" bits are effectively zero for the core entries. This is depicted in FIG. 4.
• This arrangement of core bits provides for a constant sum of the bit allocations for each of the two modes, wideband or narrowband.
• FIG. 4 shows the preferred packing of bits for the 2-way split VQ of FIG. 3.
• the configuration of the bit stream 320 comprises the mode bit 322 (indicating a wideband or narrowband input signal) followed by the "II" core bits 324 (either wideband or narrowband) and the "12" core bits. Finally the "II" narrowband extra bits or the "12" wideband extra bits complete the preferred packing configuration . Since one of the codebooks is fully searched, and for the other codebook only the core entries are searched, advantageously only one set of extra bits needs to be sent. The extra bits for the partially searched codebook will effectively be zero i.e. either 302, 306 and 308 or 310, 312, 314 need to be sent since either 304 or 316 will convey no information.
• the mode bit and core bits are beneficially always in the same locations. Hence, the impact of a mode bit error can be arranged to result in much smaller errors in the two quantisers than in prior art arrangements.
• a series of "test" input speech signals may be used, to obtain the optimum set of vectors to represent all input speech signals .
• FIG. 5 shows two octagons 350 and 354, partitioned to reflect eight separate locations identified by a respective 3-bit address.
• the two octagons 350, 354 individually each represent one of the two split VQ codebooks of FIG. 3.
• the example shows the case where three "extra" LSBs are used.
• the xxxx & yyyy represent core entry bit patterns for each of the respective embedded codebooks.
• a potentially "non-zero extra” position will be appended instead of all zeros and the "extra" LSBs of the larger codebooks will be set to zero.
• the maximum error for a (WB/NB) mode bit is equivalent to that of several LSB errors in each codebook.
• the codebook entries of the embedded codebook trained using relevant and appropriately varied speech patterns, must be interlaced regularly within the large codebook.
• index reassignment of the combined codebook must be performed such that LSB errors in the indices result in small perceptual distances. This may be arranged using a simulated annealing method as is well known to those skilled in the art.
• a set of codebooks was derived and selectively searched. In order to determine which codebook configuration to search during each frame an appropriate narrowband speech indicator was employed.
• Table 2 The results of the hybrid wideband/narrowband scheme, in accordance with a preferred embodiment of the invention, is shown in Table 2.
• Table 2 7KHz and 3KHz Spectral Distortion Results for the Hybrid 41-bit Quantiser.
• the graph 400 shown in FIG. 6 demonstrates the error resilience of the preferred embodiment of the invention in the presence of 10% bit errors, applied on a per-bit basis measured using an objective distortion measure, such as the perceptual speech quality measure (PSQM value), as defined by the ITU-T Recommendation P.861.
• PSQM value perceptual speech quality measure
• Graph 400 shows the bit error sensitivity profiles, this time for two 43-bit quantisers according to the embodiment.
• the distortion 402 (PSQM value) is shown plotted against bit number 404 on a bit-by-bit basis.
• the two quantisers shown are the hybrid 8,8,8,7,6,5 &
• the mode bit is the first bit and then the other core quantiser bits (8,8,8,7,6,2) are presented MSB first for each quantiser in-turn, followed finally by the three extra bits as depicted in FIG. 4.
• the bits are presented in MSB first natural order.
• the overall performance of the new quantiser in the presence of bit errors can be seen to be only very slightly worse than the wideband-only scheme (see rank- ordered sensitivities).
• the graph highlights that the sensitivity of the mode bit is 33rd out of 43, i.e. near, but not quite at, the bottom of the rank-ordered results.
• the explanation for the mode bit not being the least sensitive bit (as in the optimal case) positioned at the bottom of the rank ordering (see FIG. 7) is that when a mode error occurs, several LSB changes (in the three- quantiser tables) occur which together are more significant than a single LSB change (bottom of the rank ordering) . This clearly shows that the embedded structuring of the LSF VQ and bit stream has beneficially rendered the LSF VQ relatively immune to bit errors.
• the graph 450 shown in FIG. 7 demonstrates the error resilience of the preferred embodiment of the invention where the error sensitivity profile is shown rank- ordered, as compared to a bit-by-bit basis as shown in FIG. 6.
• Graph 450 shows the bit error sensitivity profiles, of the same two 43-bit quantisers, 456 and 458.
• the distortion 452 is shown plotted against re-ordered bit position 454 on a rank-ordered bit basis.
• the graph highlights that the two quantisers have broadly similar error sensitivity profiles and that the addition of the mode bit has not increased error sensitivity.
• bit-error robust embedded split vector quantiser for wideband line spectral frequencies (lsfs) in narrowband tande ming described above provides at least the following advantages.
• the invention provides for a single speech codec codebook that can quantise both wideband and narrowband signals in a near optimal manner to that of two independently- optimised speech codec codebooks. This provides for a reduced memory requirement of the codebook, in a memory constrained speech unit.
• inventive concepts described herein find particular use in speech processing units that are flexible enough to cope with a variety of bandwidth constrained speech input signals, such as future third generation cellular telecommunications systems
• any number of line spectral frequencies can be accommodated, in a LSF codebook arrangement.
• the present invention may be implemented outside of the line spectral frequency area, such as in video encoding and decoding.
• inventive concepts can benefit from such inventive concepts.
• inventive concepts can be applied to any LPC order, with any bit- division relationship.
• inventive concepts contained herein can be equally employed in any classified overlapping codebook arrangement, not necessarily limited to the overlapping arrangement between wideband and narrowband speech signals.
• bit-error robust speech codec is complementary to popular, proven speech codecs such as embedded split VQ codecs.
• the bit-error robust speech codec accommodates wideband line spectral frequencies in narrowband tandemming, and alleviates at least some of the aforementioned disadvantages.

Landscapes

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

Applications Claiming Priority (3)

Publications (2)

Family

ID=9902177

Family Applications (1)

Country Status (6)

Cited By (1)

Families Citing this family (1)

Family Cites Families (9)

• 2000
  • 2000-10-30 GB GB0026463A patent/GB2368761B/en not_active Expired - Fee Related
• 2001
  • 2001-10-22 AT AT01993000T patent/ATE303647T1/de not_active IP Right Cessation
  • 2001-10-22 WO PCT/EP2001/012403 patent/WO2002037477A1/fr active IP Right Grant
  • 2001-10-22 DE DE60113144T patent/DE60113144T2/de not_active Expired - Lifetime
  • 2001-10-22 EP EP01993000A patent/EP1334485B1/fr not_active Expired - Lifetime
  • 2001-10-22 AU AU2002215972A patent/AU2002215972A1/en not_active Abandoned

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events