Vector Quantization in Residual-Encoded Linear Prediction of Speech

M.Eng. Thesis, August 1983

Supervisor: P. Kabal

The design and implementation of vector quantizers have recently attracted considerable attention in the speech coding field. Previous work concentrated mainly upon the theoretical capabilities and asymptotic performance of vector quantizers. Little investigation concerning the actual implementation of vector quantizers was performed. It was only recently that practical algorithms have been developed for vector quantizer design.

This thesis presents an investigation into the field of vector quantization. Commencing with a review of one-dimensional quantization theory, an extension of quantization principles to several dimensions is presented. This is coupled with a survey of current work in the field of vector quantization. Based on this discussion, a vector quantizer structure, designed using the Linde-Buzo-Gray algorithm, is chosen for the block quantization of the residual signal derived from the linear prediction of speech. The performances of the residual vector quantizers are evaluated for various block sizes and transmission rates and compared to those of uniform and Lloyd-Max scalar quantizers. A subjective evaluation of residual-encoded linear predictive coders using scalar and vector quantizers is made. Finally, a subjective comparison of the linear predictive coders using vector quantization of the residual to Log-PCM coders is performed.

Sub-band Coding of Speech with Dynamic Bit Allocation

M.Eng. Thesis, August 1982

Supervisor: P. Kabal

The results of the investigation of an 8-band sub-band coder with dynamic bit allocation at the rate of 16 Kbits/sec is presented. The band-partitioning technique adopted is the Quadrature Mirror Filtering (QMF) method and particular attention is paid to the problems involved with the use of such filters in a sub-band coder.

The sub-band signals are quantized using Adaptive Pulse Code Modulation (APCM) quantizers. The number of quantization levels assigned to the sub-bands is revised regularly to adapt the coder to the changes in time of the spectral properties of speech. A simple algorithm is developed for the optimal (in the mean square error sense) assignment of the bits to the sub-bands base on the relative magnitudes of the sub-band energies.

The computer simulation of the coder produces fully intelligible speech of reasonably good quality. This study includes a discussion of the simulation of the coder as well as suggestions for improvements based on perceptual criteria.

Finally practical issues involved with the implementation of a real-time coder are considered.

Convergence Properties of Fractionally-Spaced Equalizers for Data Transmission

M.Eng. Thesis, August 1982

Supervisor: P. Kabal

This thesis considers the convergence properties of adaptive
equalizers used for data transmission. In the conventional form of a tapped
delay line equalizer, the tap spacing is equal to the symbol interval *T*.
Two other cases are discussed. The fractional-spaced equalizer has tap spacing
less than *T* (*T*/2 is considered in detail. A hybrid configuration
uses both *T* spaced and fractional *T* spaced taps is also
considered. From the mathematical derivations and the computer simulations, the
properties, the relative advantages and drawbacks of the three cases are
analysed.

Adaptive Transform Coding of Speech Signals

M.Eng. Thesis, May 1982

Supervisor: P. Kabal

Frequency domain coding techniques have recently received considerable attention. Prominent among these techniques, adaptive transform coding offers excellent speech quality for low to medium data rates (8-16 kb/sec). Adaptive transform coders divide speech into frequency components by using a suitable transform and transmit these components using pulse code modulation (PC). Three basic issues in the design of adaptive transform coders are: (1) Selection of the best transform; (2) Selection of the best quantization strategy; (3) Selection of a spectral parameterization technique.

This thesis discusses design considerations with emphasis on finding variants of adaptive transform algorithms amenable to hardware implementation. In this context coder performance using reduced frame lengths is presented. Objective and subjective performance reduction, cased by frame boundary discontinuities and low-pass filtering effects are investigated as the primary sources of perceptual distortion. Results from two computer simulations of adaptive transform coders using all-pole and homomorphic spectral fits are presented.

Telecommunication Network Modernization: A Linear Programming Formulation

M.Eng. Project, May 1981

Supervisors: L. Mason and P. Kabal

A problem encountered in telephone network planning is the choice of timing and equipment types to be used to undertake the modernization of the network. There are many options: (1) Retain network in its existing analog state; (2) Introduce digital equipment in the Switching Center; (3) Enhance analog lines with some kind of modern equipment; (4) Replace present network with an advanced digital network; (5) Combinations of the above.

In this project an aggregate model of a modernization process is presented. The economic evaluator used in this cost analysis is the Net Present Value (NPV) of cash flows which is optimized subject to capital budget constraints imposed on the modernized lines each year. Finally a linear program is used to solve for the optimal evolution strategy based on some numerical values of the input parameters.

The motivation for developing this model arose from the interest in relating modernization policies for "Wire-Center" of a local telephone network to an appropriate economic indicator and to determine the optimal policy subject to a capital budget constraint.

Codage Numérique à Taux Variable des Signaux de Parole

Master's Thesis (INRS-Telecom), April 1981

Supervisor: P. Kabal

Dans le projet qui suit, nous étudions le
comportement du codage à taux variable dans le domaine temporel. L'algorithme de
codage procède à partir d'un bloc de *N* échantillons d'entrée dont il
évalue la fonction de variance et de là, le nombre de bits à
être alloués pour chaque échantillon.

Dans la première partie du travail, nous reprenons les sections de la théorie de l'information qui traitent du problème, pour ensuite réaliser une simulation "idéale" d'un algorithme de codage à taux variable. Dans la second partie, le comportement du codeur est analysé et comparé aux prévisions théoriques. Nous terminons finalement avec un aperçu des tentatives possibles en vue de réaliser une version "pratique" d'un tel codeur.

Nous concluons de ce travail que le seul véritable gain possible dans le domaine du temps est redevable aux stries de la période fondamentale, ce qui rend impossible toute réalisation pratique d'un tel codeur puisque la transmission de l'information nécessaire relativement aux stries demande un taux d'information latérale beaucoup trop élevé. Toute tentative effectuée en vue de réduire ce taux résulte en une annulation complète du gain par rapport au codage à taux fixe.

Buffer Management for Variable-Length Encoding of Speech

M.Eng. Project, January 1981

Supervisor: P. Kabal

This report studies three methods to solve the buffer management problem introduced by using variable-length codes in digital processing of speech signals. Two feed-back coders and one feed-forward coder are proposed. The feedback coders perform adaptive quantization, variable-length encoding and buffer management. The feed forward coder performs quantization, variable-length encoding and the search for the right quantization step-size. The first feedback coder exploits the dual function of the expansion-contraction factors associated with an adaptive quantizer. These multipliers are used not only to track the change in input standard deviation, but also to adjust the bit rate into the transmission buffer. The other feedback coder controls the bit rate into the buffer by changing the quantization step-size. The feed forward coder takes a vector of samples and adjusts the quantization step-size until the total number of bits required to code this vector is approximately equal to a given value. The design, simulation, and performance of the proposed coders for speech signals transmitted over a telephone network are then discussed. Comparisons of the proposed coders and an adaptive non-uniform quantizer with fixed length encoding are also presented. The report concludes with the subjective performance of the coders and the trade-off between complexity and performance.

Thesis titles.