Robust low-complexity backward adaptive pitch predictor for low-delay speech coding

Abstract

A pitch predictor is described that is designed to operate in a low-delay vector excitation coder. The low-delay requirement of the coder necessitates a backward adaptive configuration for the pitch predictor. This adaptation is performed using a novel algorithm, called hybrid backward adaptation. As the name implies, this is a hybrid combination of backward block adaptation and backward recursive adaptation. A pitchtracking technique is described that is used to adapt the pitch period in a backward recursive manner. The pitch predictor is robust and outperforms standard backward block adaptation, while having much lower complexity. Resulting decoded speech is of very good quality, as judged by informal listening tests, and has a segmental signal-to-noise ratio (SNR) of over 20 dB. The speech degradation in the presence of increasing channel bit errors is smooth, up to a channel bit error rate of 10−3. The pitch predictor improves the SNR of the coder by about 2 dB.