Speech Recognition Performance as Measure of Speech Dereverberation Quality

摘要

Optimal, in the sense of automatic speech recognition (ASR) accuracy maximum, parameters of the late reverberation suppression technique have been proposed in this paper. It was shown that the value 50 ms as boundary between early reflections and late reverberation, which usually is used when problems of speech quality and intelligibility is studied, isn’t best for ASR systems, for which optimal value is 100 ms. It was shown also that, when estimating late reverberation power spectrum, an optimal value of averaging parameter should be associated with statistical speech constants such as phoneme and stationary durations. Several speech quality indicators were used, and it was found that recognition accuracy is the best indicator in the sense of ability to inform the user about reached compromise between reverberation suppression and speech distortion.