Susan L. Goldman
Louis D. Braida
MIT Res. Lab. of Electron., 36-749, Cambridge, MA 02139
The performance of aids for the deaf based on automatic speech recognition (ASR) is limited by errors made by the recognizer. To estimate the effects of these errors on the reception of automatically generated cued speech, acoustic speech signals were degraded to correspond to the errors made by a current ASR system. Phonemes in sentences were replaced by bursts of broadband noise according to the ASR phoneme error pattern. Five hundred and eighty sentences were generated with total segment replacement rates of 10%, 20%, 30%, and 40%. Measured phoneme error rates for auditory reception of the processed sentences (without speechreading) were substantially lower (by a factor of 4 for replacement rates of 10% to 20% and a factor of 2 for replacement rates of 30% to 40%) than the segment replacement rate. Findings are discussed as they apply to both the feasibility of using an ASR as a front end for an assistive device and phoneme restoration. [Research supported by NIH.]