Cepstral features derived from power spectrum are widely used for automatic speech recognition. Very little work, if any, has been done in speech research to explore phase-based representations. In this paper, an attempt is made to investigate the use of phase function in the analytic signal of critical-band filtered speech for deriving a representation of frequencies present in the speech signal. Results are presented which show the validity of this approach.

Speech recognition is usually regarded as a problem in the field of pattern recognition, where one first estimates the probability density function of each pattern to be recognized and then uses Bayes theorem to identify the pattern which provides the highest likelihood for the observed speech data. In this paper, we will take a different approach to this problem. In speech recognition, the goal is communication of information by voice and we will discuss the basics of speech recognition from a communication perspective. The speech signal at the acoustic level has a bit rate of 64 kb/s but the underlying sound patterns have an information rate of less than 100 b/s. What is the role of this high bit rate at the acoustic level? We will discuss the principles of decoding patterns that are submerged in an ocean of seemingly irrelevant information. 1.

The conventional n-gram language model exploits dependencies between words and their fixed-length past. This letter presents a model that represents sentences as a concatenation of variable-length sequences of units and describes an algorithm for unsupervised estimation of the model parameters. The approach is illustrated for the segmentation of sequences of letters into subword-like units. It is evaluated as a language model on a corpus of transcribed spoken sentences. Multigrams can provide a significantly lower test set perplexity than n-gram models.>

An important problem in speech coding is the quantization of linear predictive coefficients (LPCs) with the smallest possible number of bits while maintaining robustness to a large variety of speech material and transmission media. Since direct quantization of LPCs is known to be unsatisfactory, the authors consider this problem for an equivalent representation, namely, the line spectral frequencies (LSFs). To achieve an acceptable level of distortion a scalar quantizer for LSFs requires a 36 bit codebook. The authors derive a 30 bit two-quantizer scheme which achieves a performance equivalent to this scalar quantizer. The two-quantizer format consists of both a vector and a scalar quantizer such that for each input, the better quantizer is used. The vector quantizer is designed from a training set that reflects the joint density (for coding efficiency) and which ensures coverage (for robustness). The scalar quantizer plays a pivotal role in dealing better with regions of the space that are sparsely covered by its vector quantizer counterpart. A further reduction of 1 bit is obtained by formulating a new adaptation algorithm for the vector quantizer and doing a dynamic programming search for both quantizers. The method of adaptation takes advantage of the ordering of the LSFs and imposes no overhead in memory requirements. The dynamic programming search is feasible due to the ordering property. Subjective tests in a speech coder reveal that the 29 bit scheme produces equivalent perceptual quality to that when the parameters are unquantized.>

For low bit rate speech coding applications, it is important to quantize the LPC parameters accurately using as few bits as possible. Though vector quantizers are more efficient than scalar quantizers, their use for accurate quantization of linear predictive coding (LPC) information (using 24-26 bits/frame) is impeded by their prohibitively high complexity. A split vector quantization approach is used here to overcome the complexity problem. An LPC vector consisting of 10 line spectral frequencies (LSFs) is divided into two parts, and each part is quantized separately using vector quantization. Using the localized spectral sensitivity property of the LSF parameters, a weighted LSF distance measure is proposed. With this distance measure, it is shown that the split vector quantizer can quantize LPC information in 24 bits/frame with an average spectral distortion of 1 dB and less than 2% of the frames having spectral distortion greater than 2 dB. The effect of channel errors on the performance of this quantizer is also investigated and results are reported.>

The authors present an LPC (linear predictive coding) speech coder that classifies speech into periodic and nonperiodic intervals. The coder uses a stochastic codebook, as in CELP (code excited linear prediction), to synthesize nonperiodic speech and single-pulse excitation (SPE) to synthesize periodic speech. This coder is called SPE-CELP. The optimization of the single-pulse excitation is based on a new algorithm that determines the time instants of pitch periods within a short interval of periodic speech of approximately 32 ms. It thus causes a coding delay that is acceptable for many applications. Results using a fixed delta-impulse shape and time-varying pulse shapes obtained from codebooks are reported.>

Though vector quantizers are more efficient than scalar quantizers, their use for fine quantization of linear predictive coding (LPC) information (using 24-26 b/frame) is impeded due to their prohibitively high complexity. In the present work, a split vector quantization approach is used to overcome the complexity problem. The LPC vector, consisting of ten line spectral frequencies (LSFs), is divided into two parts and each part is quantized separately using vector quantization. Using the localized spectral sensitivity property of the LSF parameters, a weighted LSF distance measure is proposed. Using this distance measure, it is shown that the split vector quantizer can quantize LPC information in 24 b/frame with 1-dB average spectral distortion and>

A first-order pitch predictor is described whose delay is specified as an integer number of samples plus a fraction of a sample at the current sampling rate. This realization has a better performance than conventional multiple coefficient predictors and leads to more efficient coding of the predictor parameters. Also discussed is the application of noninteger delay pitch predictors to low-bit-rate speech coding.>

A description is presented of a framework for developing compact and accurate representation of LPC (linear predictive coding) excitation. In this representation, the excitation waveform is expressed as a linear combination of the eigenvectors of the autocorrelation matrix of the LPC filter's impulse response. This representation allows a systematic study of changes in the filter excitation on the speech output. The author presents results on the precision that is necessary in the excitation without producing perceptible distortion in the output speech signal and an estimate of the minimum number of bits necessary for accurate reproduction of the excitation.>

The authors present results on the comparative performance of nonuniform scalar quantizers using three different LPC (linear predictive coding) representations: the arcsine of reflection coefficients, the log area ratios, and the line spectral frequencies. On comparing the spectral distortion introduced by quantizers based on these representations, it was found that the average distortion was very similar for all three, with the arcsine showing fewer large spectral errors. In a parallel study, the performance of the above LPC representations and the autocorrelation coefficients for interpolating the spectrum between adjacent time frames was investigated and revealed only small differences between the different representations. Informal listening tests with a complete 8 kb/s code-excited linear predictive (CELP) coder, incorporating both quantization and interpolation, showed no significant differences between the various LPC representations, suggesting that the random codebook for the excitation is able to compensate for small spectral deviations.>

Some of the distortions produced by CELP (code-excited linear prediction) coders are characterized. It is found that the coder does not reproduce high frequencies well and that rapid changes in the speech signal are not adequately tracked. Within the framework of the current CELP concept, strategies are discussed that can reduce these distortions. Nonstationarities in the speech signal can be better followed by allowing a flexible allocation of the bits used for the excitation. However, the bit allocation procedures and the way the bits are used need further improvement. The reproduction of higher frequencies can be improved by changing the error-weighting procedure or by shaping the code-book excitation functions.>

Accurate quantization of LPC parameters with a minimum number of bits is necessary for synthesizing high quality speech at low bit rates. Earlier work by Juang and Gray has shown that vector quantization can provide a significant reduction in the bit rate needed to quantize the LPC parameters. Previous work on vector quantization of LPC parameters employed trained codebooks. In this paper, we describe a stochastic model of LPC-derived log areas that eliminates training of the codebook by constructing codebook entries from random sequences. Each vector of LPC parameters is modelled as a sample function of zero mean Gaussian stochastic process with known covariances. We generate an ensemble of Gaussian codewords with a specified distribution where the number of codewords in the ensemble is determined by the number of bits used to quantize the LPC parameter vector. The optimum codeword is selected by exhaustive search to minimize the Euclidean distance between the original and quantized parameters. Our results show that vector quantization using random codebooks can provide a SNR of 20 dB in quantizing 10 log area parameters with 28 bits/frame. An important advantage of random codebook is that they provide robust performance across different speakers and speech recording conditions.

We discuss in this paper the role of multi-pulse excitation and its importance in the synthesis of natural-sounding voiced speech. In an earlier study, Atal and David reported on the relative importance of spectral and phase properties of LPC excitation on the subjective quality of synthetic speech. This paper extends the results of Atal and David to the multi-pulse excitation. The multi-pulse analysis procedure determines the excitation by a closed-loop matching procedure which minimizes the perceptual difference between the original and synthetic speech signals. However, even for periodic voiced speech, the secondary pulses in the multi-pulse excitation do not vary systematically from one pitch period to another. We find that the multi-pulse excitation is highly periodic in the lower frequency bands and much less periodic in the higher bands. We suggest, that this irregularity contributes to the "fullness" heard in multi-pulse synthesized speech. We have investigated the influence on speech quality of replacing the original multi-pulse patterns in the excitation with a fixed multi-pulse pattern selected randomly from the multi-pulse excitation of a voiced speech segment. Our results suggest that the fixed multi-pulse patterns introduce only small degradations in synthetic speech.

Past research on CELP (Code-Excited Linear Predictive) coders has mainly concentrated on the feasibility of the CELP concept and on the reduction of the computational complexity. In this paper we address the problem of finding and encoding the excitation parameters with a limited bit rate, such that high quality speech coding in the 4.8 - 7.2 kb/s range becomes feasible. First, we examine the effect of the various excitation parameters such as code book size, code book population, order of the long-term predictor and update rate on the quality of the reconstructed speech. Second, we investigate procedures for designing and incorporating quantizers for the parameters involved. Finally, using both scalar and vector quantization techniques for the LPC coefficients, we simulated 4.8 kb/s and 7.2 kb/s coders. We also report on the use of postfiltering to further improve the performance of the CELP coder.

We will present in this paper some recent developments in low bit rate speech coding research. Two new speech coding algorithms - multi-pulse LPC and stochastic coding (code-excited linear prediction) - have been proposed recently to achieve high quality speech at bit rates below 10 kbits/sec. Both of these algorithms use a linear filter to synthesize speech at the receiver but they differ in the methods used to generate the excitation for the linear filter. The multi-pulse model assumes that the excitation can be represented with sufficient accuracy as a sequence of pulses (typically 4 to 8 pulses every 5 msec). In stochastic coders, the excitation is selected from a random codebook of white Gaussian sequences. The optimum excitation in both these coders is chosen to minimize a subjective error criterion based on properties of human auditory perception. Although these coding algorithms are complex requiring over 10 million multiply-add operations per second, new fast digital signal processor chips offer the possibility of their real-time implementation.

Stochastic linear predictive coders have the potential for producing high quality synthetic speech at bit rates as low as 4.8 kbits/sec. However, these coders are computationally very complex requiring more than 500 million multiply-add operations per second - a capability much beyond the reach of present digital hardware. Most of the complexity in stochastic coders comes from the search procedure used to determine the optimum innovation sequence from a random code book of white Gaussian sequences. This paper will describe several procedures to simplify the search in stochastic coders. The computational efficiency is achieved by moving the search from time domain to another domain where the convolution (filtering) operation in the search procedure is reduced to a single multiplication operation. The new procedures have reduced the computational load by a factor of approximately 20 without producing any additional degradation in synthetic speech.

Contamination of speech, for example by environmental noise, is sometimes unavoidable. Under such circumstances the familiar LPC analysis technique, either for low bit-rate coding or for automated recognition at the receiver, becomes fragile thus jeopardizing the system objective. In this paper we present an extended correlation matching approach for LPC analysis which results in good spectral matching between the true speech spectrum and the all-pole model spectrum, especially for the first three formants. The method has been tested on both synthetic as well as real speech and has been found to yield satisfactory results. Informal listening tests confirm the success of the technique developed.

We describe in this paper a code-excited linear predictive coder in which the optimum innovation sequence is selected from a code book of stored sequences to optimize a given fidelity criterion. Each sample of the innovation sequence is filtered sequentially through two time-varying linear recursive filters, one with a long-delay (related to pitch period) predictor in the feedback loop and the other with a short-delay predictor (related to spectral envelope) in the feedback loop. We code speech, sampled at 8 kHz, in blocks of 5-msec duration. Each block consisting of 40 samples is produced from one of 1024 possible innovation sequences. The bit rate for the innovation sequence is thus 1/4 bit per sample. We compare in this paper several different random and deterministic code books for their effectiveness in providing the optimum innovation sequence in each block. Our results indicate that a random code book has a slight speech quality advantage at low bit rates. Examples of speech produced by the above method will be played at the conference.

The multi-pulse excitation model provides a method for producing natural-sounding speech at medium to low bit rates. Multi-pulse analysis obtains the all-pole filter excitation by minimizing a spectrally-weighted mean-squared error between the original and synthetic speech signals. Although the method provides high quality speech around 10 kbits/sec, speech quality suffers if the bit rate is lowered. In this paper, we focus on problems encountered in attempting to maintain speech quality while synthesizing speech using multi-pulse excitation at lower bit rates.

This paper describes a method for efficient coding of LPC log area parameters. It is now well recognized that sample-by-sample quantization of LPC parameters is not very efficient in minimizing the bit rate needed to code these parameters. Recent methods for reducing the bit rate have used vector and segment quantization methods. Much of the past work in this area has focussed on efficient coding of LPC parameters in the context of vocoders which put a ceiling on achievable speech quality. The results from these studies cannot be directly applied to synthesis of high quality speech. This paper describes a different approach to efficient coding of log area parameters. Our aim is to determine the extent to which the bit rate of LPC parameters can be reduced without sacrificing speech quality. Speech events occur generally at non-uniformly spaced time intervals. Moreover, some speech events are slow while others are fast. Uniform sampling of speech parameters is thus not efficient. We describe a non-uniform sampling and interpolation procedure for efficient coding of log area parameters. A temporal decomposition technique is used to represent the continuous variation of these parameters as a linearly-weighted sum of a number of discrete elementary components. The location and length of each component is automatically adapted to speech events. We find that each elementary component can be coded as a very low information rate signal.

Present LPC analysis procedures assume that the input to the all-pole filter is white; the filter parameters are obtained by minimizing the mean-squared error between the filter output samples and their values obtained by linear prediction on the basis of past output samples. It is well known that these procedures often do not yield accurate filter parameters for periodic (or quasi-periodic) signals such as voiced speech. To compensate for the periodic nature of speech, an estimate of the excitation of the all-pole filter has to be made. Multi-pulse LPC obtains the best excitation for a specified bit rate by minimizing a weighted mean-squared criterion representing subjectively important differences between original and synthetic speech signals. In this paper we examine the possibility that multi-pulse excitation can approximate the all-pole filter excitation sufficiently closely and obtain the optimum filter parameters for this excitation.

The excitation for LPC speech synthesis usually consists of two separate signals - a delta-function pulse once every pitch period for voiced speech and white noise for unvoiced speech. This manner of representing excitation requires that speech segments be classified accurately into voiced and unvoiced categories and the pitch period of voiced segments be known. It is now well recognized that such a rigid idealization of the vocal excitation is often responsible for the unnatural quality associated with synthesized speech. This paper describes a new approach to the excitation problem that does not require a priori knowledge of either the voiced-unvoiced decision or the pitch period. All classes of sounds are generated by exciting the LPC filter with a sequence of pulses; the amplitudes and locations of the pulses are determined using a non-iterative analysis-by-synthesis procedure. This procedure minimizes a perceptual-distance metric representing subjectively-important differences between the waveforms of the original and the synthetic speech signals. The distance metric takes account of the finite-frequency resolution as well as the differential sensitivity of the human ear to errors in the formant and inter-formant regions of the speech spectrum.

Digital coders invariably introduce errors into analog signals such as speech. The importance of adjusting the spectrum of the error so as to minimize the subjective distortion is now well recognized. The optimum noise spectrum for minimum subjective distortion is in general non-flat. Predictive coders use instantaneous quantizers with feedback around the quantizer to shape the quantizing noise spectrum. However, for non-flat spectra, instantaneous quantizers are suboptimal. We have therefore considered and reported on non-instantaneous digital speech coders using tree-search procedures to determine the optimal innovation sequence. In this paper, we describe tree-search procedures to determine optimal block codes. We discuss both random and special block codes based on optimum packing of Voronoi regions on unit spheres in multi-dimensional spaces.

Predictive coding is a promising approach for speech coding. In this paper, we review the recent work on adaptive predictive coding of speech signals, with particular emphasis on achieving high speech quality at low bit rates (less than 10 kbits/s). Efficient prediction of the redundant structure in speech signals is obviously important for proper functioning of a predictive coder. It is equally important to ensure that the distortion in the coded speech signal be perceptually small. The subjective loudness of quantization noise depends both on the short-time spectrum of the noise and its relation to the short-time spectrum of the Speech signal. The noise in the formant regions is partially masked by the speech signal itself. This masking of quantization noise by speech signal allows one to use low bit rates while maintaining high speech quality. This paper will present generalizations of predictive coding for minimizing subjective distortion in the reconstructed speech signal at the receiver. The quantizer in predictive coders quantizes its input on a sample-by-sample basis. Such sample-by-sample (instantaneous) quantization creates difficulty in realizing an arbitrary noise spectrum, particularly at low bit rates. We will describe a new class of speech coders in this paper which could be considered to be a generalization of the predictive coder. These new coders not only allow one to realize the precise optimum noise spectrum which is crucial to achieving very low bit rates, but also represent the important first step in bridging the gap between waveform coders and vocoders without suffering from their limitations.

We describe a split-band adaptive predictive coding system for digital transmission of speech signals. In this system, the prediction residue signal obtained after spectral prediction is filtered into 2 or more frequency bands. Each of the filtered signals is reduced further by pitch prediction and is quantized by a 15-level noise feedback quantizer. The input to the quantizer is severely center-clipped to produce a quantized signal with low entropy. The division of the prediction residue signal into many frequency bands results in more accurate pitch prediction - particularly, at low frequencies. The split-band system uses separate quantizers for each frequency band. The step size of the quantizer and the center-clipping threshold can thus be adjusted to optimize speech quality in each band.

In previous papers on digital coding we have stressed the importance of taking proper account of the masking properties of the human ear in order to minimize the subjective loudness of the quantizing noise. The resulting optimal quantizing noise spectrum is in general not flat and requires the use of noise-shaping filters. This masking of the quantizing noise by the speech signal itself has allowed us to use very low bit rates (less than 1 bit/sample for the prediction residual in aa adaptive predictive coder) while maintaining high speech quality. However, if the low bit rates are realized by a (coarse) instantaneous qnantizer, the quantizing error is not white and the noise-shaping filter (in the feedback loop around the quantizer) does not produce the intended noise spectrum. In this paper, we therefore describe non-instantaneous, tree-coding methods that allow the attainment of even lower bit rates (near the theoretical rate-distortion limit) with the precise optimum noise spectrum.

Adaptive predictive coding of speech signals at bit rates lower than 10 kbits/sec often requires the use of 2-level (1 bit) quantization of the samples of the prediction residual. Such a coarse quantization of the prediction residual can produce audible quantizing noise in the reproduced speech signal at the receiver. This paper describes a new method of quantization for improving the speech quality. The improvement is obtained by center clipping the prediction residual and by fine quantization of the high-amplitude portions of the prediction residual. The threshold of center clipping is adjusted to provide encoding of the prediction residual at a specified bit rate. This method of quantization not only improves the speech quality by accurate quantization of the prediction residual when its amplitude is large but also allows encoding of the prediction residual at bit rates below 1 bit/sample.

In speech analysis and synthesis based on linear prediction, it is a common assumption that predictor coeffcients contain all the necessary spectral and phase information for accurate synthesis of the speech signal. However, even under the best circumstances, the synthetic speech sounds unnatural to the critical listener. Subjective tests reveal that spectral errors introduced by the linear prediction analysis techniques are a major source of unnatural sound quality in synthetic speech. This paper describes a modified analysis-synthesis procedure which, although relying on the basic LPC technique for analysis and synthesis, avoids spectral amplitude and phase distortions introduced by these techniques. In new method, proper reproduction of speech spectrum at the receiver is ensured by transmitting the short-time spectrum of prediction residual to the receiver.

Adaptive shaping of the quantizing noise spectrum is essential for optimal subjective performance of speech coders. In this paper, we discuss detailed procedures for minimizing the subjective loudness or audibility of quantizing noise in linear predictive coders. The procedure takes account of the frequency analysis by the human ear and its auditory masking properties.

Predictive coding methods attempt to minimize the r.m.s. error in the coded signal. However, the human ear does not perceive signal distortion on the basis of r.m.s. error regardless of its spectral shape relative to the signal spectrum. Specifically, for speech signals, the locations of the formant frequencies and their rates of change with time influence the audibility, and thus the subjective distortion of any quantizing noise. In this paper, methods for reducing the subjective distortion in predictive coders for speech siganls are described and evaluated. Improved speech quality is obtained a) by efficient removal of formant and pitch related redundant structure of speech before quantizing and b) by effective masking of the quantizer noise by the speech signal.

Recent work of Olive and Spickenagel has shown that pseudo-area parameters used for LPC synthesis can be linearly interpolated between dyad boundaries without producing excessive distortion in synthetic speech. This study investigates whether such interpolation can be done equally successfully on the power spectrum of the speech waveform. The spectrum is of special interest because speech can be synthesized in real time from spectral parameters on readily available programmable digital filters. Our results show that the distortion introduced by dyadic interpolation of spectrum is perceptually significant but it can be reduced considerably by using an additional point within the dyad boundaries for interpolation. The reasons for good quality of speech synthesized from dyadically-interpolated area parameters were also investigated. It was found that formant frequency movements are reproduced fairly accurately after dyadic interpolation. Formant bandwidths however are not reproduced accurately but the bandwidth errors are not as important subjectively.