A study of adaptive enhancement methods for improved distant speech recognition

Abstract

Automatic speech recognition systems trained on speech data recorded by microphones placed close to the speaker tend to perform poorly on speech recorded by microphones placed farther away from the speaker due to reverberation effects and background noise. I designed and implemented a variety of machine learning models to improve distant speech recognition performance by adaptively enhancing incoming speech to appear as if it was recorded in a close-talking environment, regardless of whether it was originally recorded in a close-talking or distant environment. These were evaluated by passing the enhanced speech to acoustic models trained on only close-talking speech and comparing error rates to those achieved without speech enhancement. Experiments conducted on the AMI, TIMIT and TED-LIUM datasets indicate that decreases in error rate on distant speech of up to 33% relative can be achieved by these with only minor increases (1% relative) on clean speech.