Robust spectrum quantization for LP parameter enhancement

In this paper, we investigate the denoising properties of robust vector quantization of the speech spectrum parameters in combination with a Kalman filter. The underlying assumption is that the high-energy speech regions can be used to reconstruct the low-energy regions destroyed by noise. This can be achieved through vector quantization with a properly weighted distortion measure. The performance of the proposed system, Kalman filtering with prior vector quantization, is compared with existing schemes for parameter estimation used in Kalman filtering. The results indicate significant improvement over the reference systems in both objective and subjective tests.