Neural code-excited linear prediction for low power speech compression

TLDR: This paper discusses the use of artificial neural learning methods for low bit-rate speech compression, potentially in non-stationary environments and employs two unsupervised learning algorithms: frequency-sensitive competitive learning and Kohonen's self-organizing maps.

Abstract: In this paper, we discuss the use of artificial neural learning methods for low bit-rate speech compression, potentially in non-stationary environments. Unsupervised learning algorithms are particularly well-suited for vector quantization (VQ) which is used in many speech compression applications. We discuss two unsupervised learning algorithms: frequency-sensitive competitive learning and Kohonen's self-organizing maps which have both been investigated for learning the codebook vectors in an adaptive vector quantizer. In contrast with earlier work, we have employed these learning rules in VQ of the linear predictive coding (LPC) prediction residual. The performance of these unsupervised learning algorithms in speaker-dependent and speaker-independent speech compression are presented. Our results compare favourably with those of code-excited linear prediction (CELP) requiring reduced computational power with a tolerable reduction in speech quality. We also explore the effects of limited precision on classification and learning in competitive learning algorithms for low power VLSI implementations.