We present the performance of several suboptimal algebraic quantizers in 24 dimensions. The Gaussian source is encoded at 1/2 bit per sample using the binary extended Golay code C/sub 24/ and the hexacode H/sub 6/. We also propose two new suboptimal decoding algorithms for the hexacode H/sub 6/.

This paper describes a fixed-rate quantizer using algebraic techniques as well as conventional stored codebooks to represent the wideband line spectral frequencies (LSF). It is based on two-stage split vector quantization with random and lattice codebooks in the first and second stage, respectively. We investigate the use of variable bit allocation in lattice quantization in order to capture outliers and reduce the overload distortion in spectral quantization, particularly for non-speech signals. Experimental results show that a variable bit allocation in split quantization yields performances slightly better than a fixed allocation at the same rate.

Estimation of the spectral envelope in the frequency domain allows to avoid some problems of linear prediction (LP) algorithms for voiced speech. We present a low complexity method of spectral envelope estimation from harmonics for low rate coding. The method consists in computing the harmonic amplitude spectrum using pitch-synchronous DFT with length depending on voicing, modifying this spectrum outside the telephone bandwidth to simplify modeling of the useful bandwidth and interpolating it by a frequency-domain low-pass filter. An all-pole model is then fitted to this modified smoothed version of the harmonic spectrum. The method was implemented on the harmonic-stochastic excitation (HSX) vocoder and the performance was compared with the LP algorithm similar to that used in the G.729 speech coding standard. A-B comparative tests show an important increase in perceptual quality.

This paper describes the 8 kb/s speech coding algorithm G.729 which has been standardized by ITU-T. The algorithm is based on a conjugate-structure algebraic CELP (CS-ACELP) coding technique and uses 10 ms speech frames. The codec delivers toll-quality speech (equivalent to 32 kb/s ADPCM) for most operating conditions. This paper describes the coder structure in detail and discusses the reasons behind certain design choices. A 16-b fixed-point version has been developed as part of Recommendation G.729 and a summary of the subjective test results based on a real-time implementation of this version are presented.

In this paper, we describe the enhanced full rate (EFR) speech codec that has recently been standardised for the North American TDMA digital cellular system (IS-136). The EFR codec, specified in the IS-641 standard, has been jointly developed by Nokia and University of Sherbrooke. The codec consists of 7.4 kbit/s speech (source) coding and 5.6 kbit/s channel coding (error protection) resulting in a 13.0 kbit/s gross bit-rate in the channel. Speech coding is based on the ACELP algorithm (algebraic code excited linear prediction). The codec offers speech quality close to that of wireline telephony (G.726 32 kbit/s ADPCM used as a wireline reference) and provides a substantial improvement over the quality of the current speech channel. The improved speech quality is not only achieved in error-free conditions, but also in typical cellular operating conditions including transmission errors, environmental noise, and tandeming of speech codecs.

This paper describes the GSM enhanced full rate (EFR) speech codec that has been standardised for the GSM mobile communication system. The GSM EFR codec has been jointly developed by Nokia and University of Sherbrooke. It provides speech quality at least equivalent to that of a wireline telephony reference (32 kbit/s ADPCM). The EFR codec uses 12.2 kbit/s for speech coding and 10.6 kbit/s for error protection. Speech coding is based on the ACELP algorithm (algebraic code excited linear prediction). The codec provides substantial quality improvement compared to the existing GSM full rate and half rate codecs. The old GSM codecs lack wireline quality even in error-free channel conditions, while the EFR codec provides wireline quality not only for error-free conditions but also for the most typical error conditions. With the EFR codec, wireline quality is also sustained in the presence of background noise and in tandem connections (mobile to mobile calls).

This paper describes the recently adopted ITU-T Recommendation G.729 Annex A (G.729A) for encoding speech signals at 8 kbit/s with low complexity. G.729A has been selected as the standard speech coding algorithm for multimedia digital simultaneous voice and data (DSVD). G.729A is bitstream interoperable with G.729; i.e., speech coded with G.729A can be decoded with G.729, and vice versa. As G.729, it uses the conjugate structure algebraic code excited linear prediction (CS-ACELP) algorithm with 10 ms frames. However, several algorithmic changes have been introduced into G.729 which resulted in 50% drop in its complexity, enabling a DSP implementation with a complexity of about 10-12 MIPS. This paper describes the algorithmic changes which have been introduced in order to achieve the low complexity goal while meeting the terms of reference. Subjective tests have been performed by ITU-T in both the selection phase and the characterization phase and the results showed that the performance of G.729A is equivalent to both G.729 and G.726 at 32 kbit/s in most operating conditions; however, it is slightly worse in case of three tandems and in the presence of background noise. A breakdown of the complexities of both G.729 and G.729A is given at the end of the paper.

This paper describes the enhanced full rate (EFR) speech codec that has been standardized by ETSI for the GSM mobile communications system in 1996. The codec was developed jointly by Nokia and the University of Sherbrooke. It operates at 12.2 kbit/s speech coding (source coding) bit-rate and provides speech quality equivalent to that of wireline telephony (G.726 32 kbit/s ADPCM). The algorithm is based on the algebraic code-excited linear prediction (ACELP) technology, using 20 ms speech frames. The GSM EFR speech codec provides substantial quality improvement compared to the current GSM full rate (FR) and half rate (HR) codecs. The old GSM codecs lag far behind wireline quality even in error-free conditions, while the EFR codec provides wireline quality also for the most typical error conditions. With the EFR codec, wireline quality is also sustained in the presence of background noise and in tandem connections (mobile-to-mobile calls). The codec was defined using fixed-point basic operators with complexity estimated at 18 WMOPS (below that of the GSM half-rate codec).

This paper presents an algorithm for encoding speech signals at bit rates below 4 kbit/s based on a mixed harmonic and stochastic modeling of the excitation signal. The algorithm uses robust pitch tracking and efficient voicing analysis to determine the ratios of the harmonic and stochastic components. The harmonic component is synthesized using a bank of bandpass filters while the stochastic component is synthesized using inverse STFT with overlap-and-add. Postfiltering is utilized at the decoder to enhance the quality of synthesized speech. A 2.4 kbit/s version of the algorithm was formally tested and the DAM and DRT scores showed that the coder performance is comparable to that of DoD 4.8 kbit/s Federal Standard FS-1016.

A new lattice vector quantization scheme, namely embedded algebraic vector quantizers (EAVQ), is proposed. This scheme makes use of spherical subsets of the rotated Gosset lattice RE/sub 8/ to constitute a vector-quantizer codebook. The codebook consists of several sub-codebooks and has an embedded structure. The codewords can be generated using an algebraic method and do not have to be stored. In combination with transform coded excitation (TCX) coding, this quantization technique is applied to 16 kbps wideband speech coding in order to quantize the so-called target signal in the frequency domain. Compared to a stochastic complex vector quantization scheme, EAVQ can achieve better performance and lead to significant savings of memory requirements.

This correspondence presents an algebraic vector-quantization scheme for encoding line spectral frequency (LSF) parameters used in linear predictive coding (LPC) of speech. The scheme is based on low-dimensionality regular-point lattices. The algebraic codebook need not be stored, and the optimum codevector is found through simple rounding of the input vector. Thus, the scheme results in significant savings of memory and reduced computational complexity when compared to traditional vector-quantizer solutions. The quantizer achieves an average spectral distortion of about 1 dB at 28 b/frame for the telephone bandwidth.

This paper reports on the specific contribution of the wavelet transform (WT) in the TCX coding model for audio signals. TCX, or "transform coded excitation", is a frame based coding algorithm that uses both time domain (linear prediction) and frequency domain (transform coding) approaches to exploit signal redundancies as well as frequency masking. While previous work on TCX used the discrete Fourier transform (DFT), the quality for highly non-stationary signals such as percussions was less than satisfactory. The WT has therefore been investigated as compromise between time and frequency resolution.

This paper presents an algebraic vector-quantization scheme for encoding the LSF parameters used in describing the time-varying short-term spectrum of speech in many modern vocoders. The quantizer achieves an average spectral distortion of 1 dB at 28 bits/frame for the telephone bandwidth. The scheme is based on low-dimensionality regular-point lattices. The properties of lattices are taken advantage of in both the design and the search of the quantizer codebook. Namely, this algebraic codebook need not be stored in memory and the optimum vector is found through simple rounding of the input variables instead of the usual exhaustive search. Thus, the scheme results in significant savings of memory and reduced computational complexity when compared to traditional vector-quantizer solutions.

This paper describes the application of transform coded excitation (TCX) coding to encoding wideband speech and audio signals in the bit rate range of 16 kbits/s to 32 kbits/s. The approach uses a combination of time domain (linear prediction; pitch prediction) and frequency domain (transform coding; dynamic bit allocation) techniques, and utilizes a synthesis model similar to that of linear prediction coders such as CELP. However, at the encoder, the high complexity analysis-by-synthesis technique is bypassed by directly quantizing the so-called target signal in the frequency domain. The innovative excitation is derived at the decoder by inverse filtering the quantized target signal. The algorithm is intended for applications whereby a large number of bits is available for the innovative excitation. The TCX algorithm is utilized to encode wideband speech and audio signals with a 50-7000 Hz bandwidth. Novel quantization procedures including inter-frame prediction in the frequency domain are proposed to encode the target signal. The proposed algorithm achieves very high quality for speech at 16 kbits/s, and for music at 24 kbits/s.>

A toll quality speech codec at 8 kbit/s with a 10 ms speech-frame currently under standardization by the CCITT is presented. The encoding algorithm is based on algebraic code-excited linear prediction (ACELP). Efficient pitch and codebook search strategies, along with efficient quantization procedures, have been developed to achieve toll quality encoded speech with a complexity implementable on current fixed-point DSP chips. Initial subjective tests showed that the codec quality is equivalent to that of G.726 ADPCM at 32 kbit/s in error-free conditions and it outperforms G.726 under error conditions. The codec can support a frame erasure rate up to 3% with slight degradation and performs adequately under tandeming conditions. The algorithm has been implemented on a single fixed-point DSP for the CCITT qualification test. It requires about 24 MIPS.>

The application of algebraic code excited linear prediction (ACELP) coding to wideband speech is presented. An algebraic codebook with a 20 bit address can be used without any storage requirements and, more importantly, with a very efficient search procedure which allows for real-time implementation. The authors describe an efficient procedure for searching such a large codebook deploying a focused search strategy, where less than 0.1% of the codebook is searched with performance very close to that of a full search. High-quality speech at a bit rate of 13 kbps was obtained.>

A general framework is introduced which allows both fast search and freedom in designing codebooks with good statistical properties. Several previously proposed schemes are compared from this viewpoint. A backward filtering formulation is given to show that sparse algebraic codes (SACs) (i.e., with few nonzero components) offer distinct advantages. It is shown that they reduce the optimal-search computation per codeword. They also allow control of the statistical properties of the codebook in the time and frequency domains. This control can be dynamic in the sense that it can be made to evolve as a function of the linear predictive coding model A(z). The algebraic-code excited linear prediction (ACELP) technology which allows full duplex operation on a single TMS320C25 at rates between 4.8 and 16 kb/s and which is based on SAC-driven dynamic codebooks is described.>

Error-correcting codes are considered as codebooks for high-performance vector quantization (VQ) of the IID Gaussian source at fractional bit rate. A family of good rate-one-half codes is introduced: the stretched Golay codes. The performance of these codes is compared to other good block codes, trellis-coded quantization, and other techniques. The stretched Golay codes are shown to outperform previously published results for block lengths 32, 40, 56, and 64. The good performance, together with fast decoding make these techniques attractive for applications such as low-bit-rate coding of speech.>

The authors present an algebraic code-excited linear prediction (CELP) speech coder where the innovation codebook comes from the first spherical code of the Barnes-Wall lattice in 16 dimensions. Novel fast optimal algorithms for finding the best sequence in this Barnes-Wall shell innovation codebook are described. This algebraic codebook makes it possible to design a CELP coder at 9.6 kb/s with good quality and still implementable on a current digital-signal-processing chip.>

The Leech lattice is a regular arrangement of points in 24-dimensional Euclidean space that yields an extremely dense packing when equal spheres are centered at these points. A subset of the Leech lattice can be used as a signal set for the Gaussian channel or as representative vectors for a vector quantizer. Of particular interest are the spherical codes (or code books) that consist of the points of the Leech lattice which lie on a sphere centered at the origin. The code points do not have to be stored because they can be obtained from a very small set of basic vectors using permutations of the components in a manner dictated by the words of the extended Golay code. A nearest-neighbor algorithm that works on this is developed to determine the point in the code closet to some arbitrary vector in R/sup 24/. The performance of this approach when quantizing independent identically distributed Gaussian samples is reported.>

The paper discusses how binary error-correcting codes can be used to provide (+1/-1)-waveform codebooks that speed up search in CELP vocoders. Four coding techniques operating at half bit per sample with respectively 8, 16, 24 and 32 samples are compared in terms of complexity and SNR performance, Recent results on spherical codes from regular point lattices are also reported.

Code-Excited Linear Prediction (CELP) produces high quality synthetic speech at low bit rate. However the basic scheme leads to huge computational loads. The paper describes a related scheme, which allows real time implementation on current DSP chips. The very efficient search procedure in the codebook is achieved by means of a new technique called "backward filtering" and the use of algebraic codes. RSB performances are reported for a variety of conditions.

The extended Nordstrom-Robinson code is an optimum nonlinear double-error-correcting code of length16with considerable practical importance. This code is also useful as a rate1/2vector quantizer for random waveforms such as speech linear-predictive-coding (LPC) residual. A fast decoding algorithm is described for maximum likelihood decoding (or nearest neighbor search in the squared-error sense) with304additions and128comparisons.

This paper concerns the design of a speech/data discriminator with the speed and accuracy required for statistical multiplexing of speech and high-speed baseband data on a single telephone line. The structure proposed for the discriminator combines the information from two parallel processes. The first process, even though slow, accurately recognizes the speech/data nature of the signal. This nature is determined by statistical pattern classification applied to simple zero-crossing-type parameters extracted from the signal filtered in a PCM representation. The second process has temporal accuracy in detecting transition events, from speech to data and vice versa. This detection is provided by special parameters which supply tentative transition markers. This paper applies these processes to the case of 9600 bit/s modems and describes a complete discrimination algorithm which is fit for microprocessor technology. The simulation results provided indicate a very low error rate and demonstrate that transitions are detected exactly or within one or two sampling intervals.

This paper shows a technique of encoding the LPC residual which allows the achievement of speech coding with residual excitation at a bit rate as low as 2400 bps. The method is inspired by the multipulse coding approach introduced by Atal, associated with an irregular downsampl-ing. The real time implementation of a 4800 bps vocoder on a single TMS 320 DSP is discussed.

This paper concerns a new Quantization Scheme for efficient encoding of waveforms below or about one bit per sample. This technique is then applied to the encoding of baseband residual signals to demonstrate the feasibility of (baseband) Residual Excited LPC at 2400 hit/sec. The technique called "Spherical Vector Quantization" is described in which a block of n consecutive samples is quantized as a vector. The magnitude of the vector is transmitted independently of the vector's orientation. This vector's orientation is vector quantized using a codebook which can be seen as representing a set of N points on a unit hypersphere. The cases for blocks of n = 8 and 24 are discussed which make use of results by Conway and Sloane on regular point lattices. For n = 8, the algorithm is detailed which solves both the problem of finding the closest point on the hypersphere and the problem of determining the index of that point. Experiments involving SVQ in coding the speech baseband residual are described which show in particular that subband coding does not contribute any quality improvement when SVQ is used.

For digital modulations applied to the coding of speech signals, a fixed sampling and transmission rate is always chosen. For commercial telephone these rates are respectively, 8 KHz and 64 Kbits/sec, corresponding to a filtered signal bandwidth of 300 to 3300 Hz. A new processing concept (variable sampling and digital quantization) is proposed where a sample is coded with a single binary word. The code word corresponds to an information pair: a variable and adaptive sampling time and a coding angle associated to the signal. The basic principle is conceived around a distribution of the coded samples in amplitude and time (amplitude-time coding) along an adaptive coding curve associated to each coded/decoded sample of the original signal. A variable sampling rate requires a buffer, thus a delay, for transmission at a fixed rate. The transmitted signal so obtained is a transposition of the original speech signal and consequently its characteristics (bandwith, amplitude dynamic range) are modified. Some of the characteristics of the transmitted signal are ultimately used for the digital or even the analog transmission of the signal.

The paper presents the full description and discusses the performances of a 4800 bit per second residual excited linear prediction vocoder. The LPC analysis is efficiently performed using a type of binary-tree search vector-quantization approach. The technique, which is described in ref (1), uses a set of hyperplane equations to perform a hierarchical pattern classification of the input autocorrelation vector in the autocorrelation space. The end result of the search is the integer i1which is the index of the most appropriate (in the Itakura-distance sense) prediction filter out of a set of N preset filters. The search requires only\Log_{2}Ndot products. In this case vector quantization presents two advantages over the classical approach of the Durbin algorithm followed by scalar quantization. First, a faster algorithm is obtained. Second, the same accuracy in filter representation is possible with less bits per second and consequently more bits can be allocated for representing the residual and gain. The residual is vector quantized in the time domain by blocks of 16 the samples according to the approach of ref (2). The 16 sample block is essentially encoded using the integer I2which is the index of the most appropriate 16-sample waveform out of set of M preset prototype waveforms stored in memory. The paper includes preference testings for comparison with other types of 4800 Kbit/sec vocoders. Some sample recordings will be presented at the conference. Finally, preliminary results in the attempt to implement the vocoder in real time on a MAP 200 array processor are discussed.

This paper presents a technique for reducing the bit rate of μ-law PCM-coded speech from 8 bits/sample to a format of 6, 5, or 4 bits/sample. The particular format can be specified dynamically, i.e., it can be made to change from sample to sample. A device based on this technique has been designed using a 3 kbyte ROM and an uncommitted logic array (ULA) circuit to accommodate up to 96 multiplexed channels. The device can be set to operate in either the bit-rate-reducing mode, for removing the redundancy at the transmitter, or in the bit-rate-restoring mode, for recovering the standard PCM speech representation. The device finds useful applications in the context of multichannel telephony. In particular, it can be combined with digital speech interpolation to yield a three to four compression factor [2]. This technique provides approximately equivalent performances to the nearly instantaneous companding (NIC) technique [1] without the need for block processing, delay, or transmission of side information. The principle is based on the use of switched optimum reencoders (quantizers) selected by using a backward adaptation scheme. It further incorporates a redundancy reduction scheme [2], [3] which capitalizes on properties of both companded PCM and speech

The paper presents a new approach to the problem of speech coding in the range of 4 to 8 kbits/sec along with experimental tests performed on a MAP 200 array processor. The basic idea consists of using a dictionary of K indexed waveforms of fixed duration T (where T lies between 1 and 4 ms). The input signal is broken down into blocks of duration T. For each segment the Dictionary is searched for that prototype waveform which is, with respect to some distance measure, the closest to the input block. The digital representation of the signal is thereafter the index of the prototype. The paper addresses the problem of the judicious choice of distance measures and the design of a good dictionary. Experimental results are provided for dictionary coding of the signal using a distance measure combining time and frequency domain features. Experimental results are further presented for the coding of the LPC residual.

This paper explores the intermediate solutions between fixed prediction and forward adaptative prediction in ADPCM which consists of using a finite number, L, of preselected linear predictors of order M. The design problem of selecting the optimum set of predictors with respect to the overall prediction gain is formulated and an iterative procedure is described to obtain the solutions. The relative prediction-gain improvement is computed for a 3 sec. speech sample and for several values of L,M, and block size showing that\frac{1}{2}of the adaptative over fixed-prediction improvement in dB is reached with onlyL=4and 2/3 withL=8. The design problem solved by minimizing Itakura distance is shown to yield essentially identical performances. A linear discriminant property in the autocorrelation space is pointed out. Based on that property a pattern classification approach is proposed as an hardware-efficient coding algorithm.

The paper present an original numerical algorithm for reducing by a factor of two (32 Kbits/ sec) the bit rate of telephone speech already digitized in standard PCM. At the end of the line (i.e. expensive media) an inverse algorithm resynthesizes a PCM signal which preserves a high telephone quality. This bit-rate compression algorithm is thus essentially "transparent" to PCM transmissions and therefore finds immediate applications for increasing carrier capacity. In particular when used in connection with speech interpolation techniques [1] it enables a fourfold increase in the number of channels per carrier.

This paper presents a class of digital transformations which generaten-level sequences having an autocorrelation function which remains nonzero throughout its domain in order to insure a narrow spectrum. In addition, any transformation of the class attains maximum redundancy and can be implemented by a simple one-bit memory stochastic machine. Most importantly, the autocorrelation function is obtained analytically which permits establishing for any odd value ofnthe set of level values which achieve optimum bandwidth compression.

This correspondence is concerned with binary processes and presents results with immediate applications in the modeling of digital channels for the purpose of evaluating code performance. It is demonstrated that cluster density can be analytically described from the distributions of intervals between errors. These relations and derived clustering properties hold for any stationary process. Analyses of real error data exemplify the use of these results in regard to channels having dependent inter-error intervals.

We present a new descriptive statistic for channels with memory and show its utility a) in evaluating and comparing existing models for such channels and b) as a theoretical tool in defining the error-gap distribution characteristics of real channels. We demonstrate that certain kinds of real channel behavior cannot be adequately described by previously proposed models and offer an example of a better model that includes many of the earlier models as special cases.