Adaptive post-filtering of ADPCM speech

We show that the quality of adaptive differential PCM (ADPCM) speech can be significantly improved by passing it through a reconstruction low-pass filter that is matched to an appropriately defined short-time speech cutoff frequency. Practically, the adaptive procedure involves switching the decoder output into one of a bank of N low-pass filters whose cutoff frequencies span the expected range of input speech bandwidth. For the case of equally spaced filter cutoffs, and with uniform probability density function models for the quantization noise spectrum and the cutoff frequency, more than one-half of the maximum adaptive filtering gain is realisable by a bank of four filters. Computer simulations of l6- and 24-kilobit/s ADPCM coders indicate that perceived quality gains are in fact greater than what is indicated by an analytically predicted objective gain of 2 6 dB.