Improving predictive vector quantizers in speech coding applications
J. Linden, Jan Skoglund, Thomas Eriksson · 1996
In this paper, we study the intervector correlations of some parameters often employed in speech coding. The speech parameters show a switched behavior, with high correlation for most vectors, but low intervector correlation for a few. We demonstrate that a quantization scheme that takes the switched behavior into account can outperform traditional methods. 1. MOTIVATION In the source-filter model of speech, the speech signal is modeled as the response of a time-varying linear filter to a voice source excitation signal. The production filter is often estimated by linear predictive analysis (LPC), where the parameters of the filter are chosen to minimize a quadratic error criterion. In speech coding applications, the excitation and filter parameters are quantized and transmitted to the receiver, and the received parameters are then used to produce synthetic speech. Figure 1 illustrates the source-filter model for speech production. a Excitation signal Synthetic speech SOURCE FILT...