No Residual Transmission: Joint Spectral-Residual Quantization

Abstract

In this chapter, we present a unit-selection based segment quantization scheme which leads to the interesting possibility of not having to transmit any side-information about the residual at all. We propose such a ‘no residual transmission’ scheme in both the segmental unit-selection framework of Lee and Cox described in Chap. 3 and the optimal 1-pass DP based unit-selection framework proposed by us and described in Chap. 4. We arrive at this ‘no residue transmission’ scheme from the important observations that unit-selection based segment quantization systems typically employ large unit-databases as in concatenative speech synthesis and that, by virtue of the largeness of the continuous codebook, it becomes possible to quantize an input segment by an unit in the unit database in such a way that the speech corresponding to the unit, after applying ‘only’ duration modification, is a close reconstruction of the input speech (of that input segment). We propose a ‘joint spectral-residual quantization’ by defining various ‘composite measures’ that combine both the spectral match and residual match, between input speech frames and unit frames, to select units that quantize an input speech segment/frame in toto. We present the rate-distortion performance trends of such a joint spectral-residual quantization in both these unit-selection frameworks for various composite measures, and show the efficacy of the proposed approach.