Evaluation of adaptive speech coders under noisy channel conditions

An experiment has been performed in which the digital transmission of speech coded by adaptive differential pcm was simulated under noisy channel conditions. The experiment was done with two aims: (I) to get information on the subjective effect of channel errors and the influence of various design parameters on the speech quality under various conditions and (ii) to find objective measures for predicting the overall quality of the processed speech over a wide range of circuit conditions. The subjective results show that, for a speech transmission through a channel with bit error probability up to 1/256, best results can be obtained with a slow error recovery, associated with fast quantizer adaptation. The use of slow error recovery and slow quantizer adaptation is preferable for channels with very high bit error rates, like 1/32. Overall subjective quality is well predicted by the sum of two terms: (I) an objective performance measure of the noise present on the output signal, disregarding any effect of level mismatching due to the sensitivity of the adaptation algorithms to channel errors and (ii) a measure of the level mismatching which takes into account both the average gain on the output signal and its fluctuation in time. The best prediction scores are achieved by three newly defined objective performance measures, two-level compensated segmental snrs, and a spectral signal-to-distortion ratio.