Fast VQ of multi-tap pitch predictor coefficients

Abstract

Pitch predictors are important in achieving high-quality speech for linear prediction-based analysis-by-synthesis (LPAS) coders. Multi-tap pitch predictors with vector quantization (VQ) of the predictor coefficients have been adopted in more and more LPAS coders because they can provide high prediction gain. Higher-tap pitch predictors (tap>3) have better performance but they are seldom used due to the high computational complexity of vector quantizing the pitch predictor coefficients. This paper proposes a new VQ scheme for multi-tap pitch predictors, the likelihood error criterion-based vector quantization (LEC-VQ). Experiments show that the proposed VQ method can efficiently reduce the computational complexity while maintaining high speech quality.