A report is presented on experiments using a colored-noise assumption Kalman filter to enhance speech additively contaminated by colored noise, such as helicopter noise and jeep noise, with a particular application to linear predictive coding (LPC) of noisy speech. The results indicate that the colored-noise Kalman filter provides a significant gain in SNR, a clear improvement in the sound spectrogram, and an audible improvement in output speech quality. The authors demonstrate that such gains are unavailable with white noise assumption Kalman and Wiener filters. The colored-noise prefilter greatly enhances the quality and intelligibility of LPC output speech for noisy inputs.>

The block and alphabet-constrained formulations are compared for the problem of vector quantization of noisy speech. In the optimum estimator/source-coder structures, a training mode vector quantizer due to Linde et al. (1980) is used as the source coder for the estimator outputs in all cases, and three block estimators and five alphabet-constrained estimators are examined. Objective and subjective performance results are obtained for all eight estimators used in conjunction with training mode vector quantizers at a rate of 1 bit/dimension, for dimensions 1,2,. . . ,8, and 2 bits/dimension, for dimensions 1,2,3, and 4, on five sentences, The results show the superiority of the alphabet-constrained approach, using the frame-adaptive Kalman filter, with improvements in output signal-to-noise ratio over the block approach of 20%.>

Simulation results are presented which compare the performance of all-pole, all-zero, and pole-zero predictors in ADPCM at data rates of 16 and 32 kbits/s over both ideal and noisy channels. Separate backward adaptive gradient algorithms are used to adapt the poles and the zeros independently. The performance indicators used are signal-to-quantization noise ratio (SNR), signal-to-prediction error ratio (SPER), segmental SNR (SNRSEG), and subjective listening tests. For speech sources, the all-zero and pole-zero predictors produce SNR and SNRSEG values that are approximately 1-3 dB higher than those generated by the all-pole predictor. Subjective listening tests reveal that an eighth-order all-zero predictor performs as well or better than an allpole predictor for all conditions studied.

A modified differential encoding structure is proposed and optimized based upon the concept of preposterior analysis from the theory of alphabet-constrained data compression. Using preposterior analysis, the quantizer input sequence is chosen to minimize the expected distortion over a fixed but arbitrary interval, say,Nsamples long. By computing the expected distortion over future inputs, preposterior analysis allows future behavior to be modeled, but without an encoding delay as in tree coding. The optimized quantizer input sequence is not simply the prediction error as in classical differential pulse code modulation, but it is a weighted combination of the current prediction error and past encoding errors. The optimization is accomplished using a backward dynamic programming argument.