Reverberation Mitigation of Speech

Automatic speech recognition (ASR) technology is in wide use today and has been successfully integrated into a number of applications. Although there are also many potential applications for automatic speaker identification (SID), SID accuracy rates are generally not sufficient for widespread use in practical applications. For high accuracy ASR and SID, the microphone must be located near the talker. However, when the speaker is at some distance from the microphone as in a hands-free communication, or in a meeting room, there is a major problem with the captured speech signals caused by interfering sounds and reverberations. While traveling from the speaker to the microphone, the wavefront of the speech is repeatedly reflected off the walls or other objects in the room. These reflections, perceived as reverberation, alter the acoustic characteristics of the original speech. The SID recognition accuracy of reverberant speech degrades considerably. During Phase I, we demonstrated ~5% Equal Error Rate (EER) performance for SID using artificially reverberated telephone speech from the NTIMIT and Mixer-6 databases. A continuation of this work will result in even more powerful solutions for speech reverberation mitigation.A high accuracy SID system will thus be created in Phase II for reverberant speech.