Analysis-by-Synthesis Speech Coding Based on Relaxed Waveform-MatchingConstraints

摘要

Methods to encode the speech signal at bit rates of around 3000 to 5000bits/second (b/s) are described. The current status of low bit rate speech coding techniques is discussed with emphasis on issues relevant for the understanding of a class of coders which are referred to as Linear Prediction (LP) based analysis by synthesis coders. Such speech coders are now capable of reconstructing good quality speech from a bit stream with a rate of approximately 5000 b/s. Special fast procedures the CELP (Code Excited LP) algorithm were developed to make implementation on modest hardware possible. A unified description of fast procedures described in the literature and a comparison of their computational and storage requirements is given. The concepts which led to the interpolation of prototype waveforms can also be used to obtain significant improvement in efficiency of LP based analysis by synthesis methods such as CELP. Methods to perform proper interpolation of the LTP parameters are provided. The procedures described have in common a relaxation of the constraints used in existing LP based analysis by synthesis speech coders, which does not affect the perceptual quality of the reconstructed speech. As a result, a speech signal of given perceptual quality can be encoded at lower bit rates.