Switched prediction and quantization of LSP frequencies
H. Zarrinkoub, Paul G. Mermelstein · 2002
We present new results on switched prediction and quantization of the line spectral frequencies applicable to high-quality coding of speech at rates below 8 kb/s. The predictor-quantizer exploits both the temporal inter-frame correlations among the line-spectral frequency vectors of successive frames as well as the spatial within-frame correlations. Best results are obtained by separately training two split vector quantizers, one on predictable frames and one on non-predictable frames. Line-frequency differences are found to yield the lowest quantization error for both the high and low order quantizers for the non-predictable condition, as well as the low order quantizer of the predictable condition. With 10 ms frame separation, 18 bit split VQs, one for each mode, with 1-bit mode indication yield a mean spectral distortion near 1 dB. With 20 ms frame separation, 20 bit quantizers are required.