DESIGN OF A 16 kb/s ADPCM CODER.

Speech coder design has traditionally been accomplished by optimizing the long-term signal-to-noise ratio, even though this measure is known to be an inadequate index of coder quality. This memorandum describes the design of a 16 kb/s ADPCM coder by the optimization of a different quality estimate. The estimate is a linear combination of two separate measures of overload distortion and granular noise. In a recent experiment it has been found that this quality estimate is very closely correlated with the rating that the average listener assigns to speech processed by ADPCM coders. The adaptive quantizer in the ADPCM coder is formulated as an estimator of the mean absolute value of the quantizer output. With this formulations the relationship of the performance to design parameters can be readily described. The adaptation mechanism however is identical to the well known one-word memory scheme and the optimum step-size multipliers for this scheme are also derived from optimum signal-estimator design parameters.