Application of exponential power estimator for speech coding

Abstract

In speech coding, it is known that the Differential Pulse Code Modulation (DPCM) schemes lead to the reduction of bandwidth requirement (by removing the redundancy) as compared to pulse Code Modulation (PCM) schemes, for a specified speech quality. To achieve adequate dynamic range and good subjective quality, however, it is necessary to adapt the parameters of the quantizer and predictor to the input signal characteristics. Several algorithms have been developed for adapting the quantizer and predictor, most of which suffer from high computational complexity and are not easily implementable. The study reported in this thesis is concerned with the development of adaptive PCM and DPCM (APCM and ADPCM) systems which are based on computationally simple adaptive quantization and prediction algorithms. Two points are noteworthy: firstly, variance estimation of the quantizer input forms the basis of adaptive quantization and secondly, in predictor adaptation using the Least Mean Square gradient (LMS) algorithm it is the estimation of the reconstructed signal power which is important. Hence, it is possible to reduce the computational complexity of the APCM and ADPCM systems by resorting to computationally simple power estimation schemes. This is achieved here by implementing the adaptation algorithms for the quantizer and predictor using the Exponential power Estimation Technique (EPE), which is a modified form of an existing power estimation method. The estimation using EPE avoids multiplications and divisions and it expresses the power estimate in the form 2^, where i is an integer. Hence the use of EPE leads to a significant reduction in computational complexity. Comparative studies of EPE with two conventional power SYNOPSIS ii estimation schemes (Block and Recursive power estimators: BPE and RPE) have been presented in the context of step response, spectral parameter identification and adaptive prediction of speech. The results indicate that EPE is much superior to RPE and identical to BPE in performance. However,' the performance of EPE is limited by the finite number of distinct power estimate values which the EPE estimate is allowed to take. The adaptive quantizer used in the APCM and ADPCM systems is of the hybrid type. EPE is used to estimate the instantaneous and syllabic standard deviations of the quantizer input signal, for the purpose of adaptation. In order to make the quantizer robust to transmission errors, a leakage mechanism has been introduced in the EPE algorithm, performance evaluation study of these systems have been done under noise free and noisy channel conditions using uniform and nonuniform quantizing characteristics at different bit rates. The input signals used are sinusoidal, correlated Gaussian sequence and digitized speech. segmented signal to noise ratio is the performance measure used. The APCM systems simulated at bit rates 24 and 32 kbps using the EPE-quantizer (EPEQ) is found to provide almost identical performance as that of jayant's adaptive quantizer for speech input while the performance degrades for sinusoidal input (due to the speech specific design of EPEQ). A detailed performance evaluation of the ADPCM system, whose quantizer and predictor are adapted based on EPE, led to the conclusion that the EPE based ADPCM is best suited for low bit rate applications. At 16 kbps, this system outperformed the two conventional ADPCM systems considered for comparison purpose, with the computational advantage of having (3N + B + 1) multiplications and one division compared to the first system and (B + 1) multiplications compared to the second system, where N is the predictor order and B is the ill number of bits/sample. However, EPE requires one extra exponentiation operation. Further, modifications are done to improve the quantizer performance. These include introduction of additional step sizes in the quantizer and modifications aimed at improving the dynamic range and noisy channel performances. The thesis has been organized in the following manner: The significance of a computationally simple power estimation scheme in the context of adaptive quantization and prediction is brought out in chapter 1. This chapter also presents a review of those speech digitization methods which are relevant to the topic of this thesis. A detailed introduction and study of the Exponential power Estimation technique is given in chapter 2. The application of EPE in adaptive quantization is dealt with in chapter 3. In Chapter 4, the development of a computationally simple ADPCM system (16 kbps) which makes use of EPE for adapting its quantizer and predictor is considered. The results are summarized in chapter 5. The thesis concludes with some suggestions for further work in this- area.