A shaped trellis coded quantizer for sampled speech

Abstract

Due to recent developments in digital communication technology there is increasing need to achieve high quality speech at low bit rates. Most speech algorithms based on VSELP and other similar systems are able to achieve the required quality, but are unable to provide low delays. In order to meet with the low delay requirement, there is a need for more efficient speech coders. In this work, a new 16 kbit/sec low delay speech coder is developed. This coder, called the "S-TCQ speech coder" uses an efficient backward adaptive scheme in order to provide excellent quality speech. The inherent quantization scheme in the S-TCQ speech coder, namely, shaped trellis coded quan.tizatz'on was developed recently by Larola et al. [1]. They have demonstrated that the S-TCQ quantizer achieves high granular gains (due to trellis coding) and high shaping gains (due to SVQ shaping). In this work, as a first step towards the development of a speech coder, the shaped trellis coded quantization scheme has been studied on a Gauss-Markov source which serves as a good model for sampled speech. The performance of this scheme has been found to be better than most other source coding schemes for Gauss-Markov sources. This powerful quantization scheme has then been applied for the coding of sampled speech. In addition, a high order short-term predictor and a variance estimator are used to exploit the redundancies in speech waveforms. In order to provide low delay a backward block adaptive scheme is used. Simulation results for this coder show excellent performance. This coder compares favorably with other existing 16 kbit/sec speech coders. SNR values in the range 16.0-19.5 dB are obtained for standard test speech files. Further, a generalized speech coder using noz'se feedback codz'ng is developed. This technique incorporates frequency-weighting of noise. Smaller block length is used in order to achieve an overall delay of 5 ms for this coder. Simulation results show a noticeable improvement in the perceptual quality. The encoded speech has little background noise and is natural sounding.