Evaluation of Adaptive Speech Coders Under Noisy Channel Conditions

The design of digital waveform coders for speech communications must face the inevitable presence of channel errors. Adaptive coders, like ADPCM (adaptive differential PCM), in which the adaptation of the 1369 quantizer step-size is derived from the transmitted binary stream and no error-protected side information is sent to the receiver, may be particularly sensitive to this problem. In fact, a single channel error may cause a multiplicative offset between the signal level at the receiver and that at the transmitter. This offset may persist indefinitely if no error dissipation mechanism is provided. Recently, some algorithms of quantizer adaptation have been developed that make the effect of a single transmission error die out over time, so the transmitter and receiver can resynchronize their step-size estimates. 1,2 The possibility of obtaining such results is physically due to the fact that these algorithms have an imperfect adaptation: the step size increases more quickly and decreases more slowly for low input levels than for high ones. In this way, the step size is overestimated for low input levels and underestimated for high ones, thereby reducing the dynamic range of the coder. Dynamic range and error dissipation rate vary inversely, and the designer has to balance between them. In the case of speech transmission, the choice of the appropriate design parameters must be based on a precise evaluation of the subjective quality of the coded speech. Use of the conventional longterm signal-to-noise ratio as an estimator of the subjective quality would be, in this instance, completely misleading, at least because an offset in the signal amplitude between input and output, due to an offset in step sizes caused by an error, will be reflected in a noticeable squared difference between the two waveforms, while it may not be subjectively disturbing.