Model-based multirate representation of speech signals and its application to recovery of missing speech packets

摘要

When the samples of a critically sampled speech signal are lost, objectionable aliasing occurs and perfect recovery of the original speech becomes impossible. In this work, a multirate state-space representation of the autoregressive (AR) speech process is derived to describe the generation of regularly missing-sample speech sequences. Next, a new sample-interpolation algorithm based on the multirate Kalman reconstruction filter is proposed to reduce speech quality degradation caused by packet losses. This method is used together with packet interleaving configuration, thereby simplifying the recovery of missing packets to the interpolation of missing samples. Subjective tests indicate that the proposed Kalman-based sample-interpolation algorithm performs better than the conventional odd-even sample-interpolation procedure for mitigating the effects of random packet losses in 64 kb/s PCM codes. The tolerable packet loss rate P/sub L/, which is strictly input-speech-dependent, can be as high as 10-20% with Kalman interpolation. These observations are based on computer simulations in terms of signal-to-noise ratio (SNR) values, waveform reconstruction plots, error spectral shapes, and summaries of informal listening tests.