Shape-gain matrix quantizers for LPC speech

Abstract

It has been recently demonstrated that the principles of vector quantization for LPC speech can be simply extended to encompass matrices of LPC vectors with significant savings in bit rate. Unfortunately, however, such locally optimal matrix quantizers have prohibitively high complexity and memory requirements when implemented in a speech vocoder at bit rates giving acceptable quality speech. One approach to solving the problem is to separately code gain and shape in the matrix quantizer. This paper generalizes the principles of shape-gain vector quantizer design for LPC speech to matrix quantization and investigates the properties of the resulting quantizers. In particular, we present a design which combines shape matrices consisting of N shape vectors with K-dimensional gain vectors, where N and K are small integers, in practice, with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K \geq N</tex> . Experimental results show that with K, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N \geq 3</tex> , significant reductions in bit rate over locally optimal vector quantizers are obtained for comparable performance. Simulations indicate that a shape-gain matrix quantizer, using a 10 bit shape codebook and an 8 bit codebook with K = N = 3 operating at 6 bits/frame for the LPC model, gives speech quality comparable to a locally optimal vector quantizer at 9 bits/frame. The matrix quantizer has somewhat greater than 5.7 times the memory requirement of the above vector quantizer, but less than 2.1 times the complexity. Subjective tests show that the speech from this matrix quantizer is intelligible to native speakers of English.