Backward Adaptive Configurations for Low-Delay Vector Excitation Coding

Abstract

Many of the advances in speech coding in the past decade at rates of 4.8–16 kbit/s have been based on excitation coding by means of analysis-by-synthesis. Excitation coding schemes have a decoder structure consisting of an excitation signal applied to a time-varying synthesis filter to produce the reconstructed, or “synthesized,” output speech. In addition to other tasks, the encoder must determine a suitable excitation signal and transmit data that specifies this excitation. In the analysis-by-synthesis technique, the excitation is selected by a closed-loop search procedure where a candidate excitation signal segment is applied to the synthesis filter, the synthesized waveform is compared with the original speech segment, the distortion is measured, and the process is repeated for all excitation segments stored in an excitation codebook. The index of the “best” excitation segment is transmitted to the decoder, which retrieves the excitation segment from a codebook identical to that at the encoder. The parameters of the synthesis filter are computed using well-known linear prediction analysis techniques on a frame of buffered input samples and transmitted to the decoder. This coding scheme is often called Vector Excitation Coding (VXC) or Code Excited Linear Prediction (CELP) [1,2].