Efficient waveform coding of speech using vector quantization

Abstract

A vector quantizer operates on blocks (vectors) of k contiguous samples of the input signal, rather than on isolated samples as in scalar quantization. Rate-distortion theory guarantees a performance arbitrarily close to optimal for waveform vector quantizers if the dimension of the vector, k, is sufficiently large. The main obstacle in using this result is implementational complexity. A vector quantizer of dimension k operating at a rate of r bits/component requires a number of computations of the order of k2('kr) and a memory of the same order. Various suboptimal vector quantization techniques, such as tree structured codebooks or multi-stage quantizers, achieve an important reduction in computational complexity. The resulting quantizers compare favorably with scalar quantization, but the achieved overall performance is not competitive with sophisticated scalar coding systems using compression techniques such as transform coding, linear prediction, and adaptive methods. A different approach for achieving a viable waveform coding system is to exploit the performance of a low dimensionality vector quantizer by combining it with other waveform compression techniques. In this approach the vector quantizer becomes a building block in the waveform coding system. Consequently, the entire system must be designed as a vector processing system as a result of the presence of the vector quantizer. The main goal of the study is to present and analyze a waveform coding system in which a low dimensionality vector quantizer is used in an adaptive differential coding scheme. In the encoding process, a locally generated prediction of the current input vector is subtracted from the current vector and the error vector is coded by a vector quantizer. Each frame, consisting of many vectors, is classified into one of m statistical types. This classification determines which one of m fixed predictors and of m vector quantizers will be used for coding the current frame. This system is called Adaptive Differential Vector Coding (ADVC). The design methods for ADVC are developed for two different fidelity criteria: the standard mean square error and the energy weighted mean square error. The ADVC system was designed and tested by simulations for the encoding of speech waveforms. A typical simulation result for the ADVC system is a signal to noise ratio slightly larger than 20 dB at a rate of 2 bits/sample. These results were obtained inside a training sequence represented by a speech file with 6 different talkers, four male and two female, totaling about 16 sec. of speech, sampled at 8 KHz (130,000 samples). This performance is competitive with the best known results for waveform speech coding including such high complexity methods as tree and trellis encoding. . . . (Author's abstract exceeds stipulated maximum length. Discontinued here with permission of author.) UMI