Robust hands-free Automatic Speech Recognition for human-machine interaction

摘要

In enclosed environments where robots are deployed, the observed speech signal is smeared due to reverberation. This degrades the performance of the automatic speech recognition (ASR). Thus, hands-free speech recognition for human-machine communication is a difficult task. Most speech enhancement techniques used to address this problem enhance the contaminated waveform independent from that of the ASR. However, this approach does not necessarily improve ASR performance. In this paper, we expand the conventional spectral subtraction-based (SS) technique to deal with reverberation. In our proposed approach, the dereverberation parameters of SS are optimized to improve the likelihood of the acoustic model and not just the waveform signal. The system is capable of adaptively fine-tuning these parameters jointly with acoustic model training for effective use in ASR application. We have experimented using real reverberant data collected from an operational robot. Moreover, we also evaluated with reverberant data corrupted with environmental and robot internal noise. Experimental results show that the proposed method significantly improves the recognition performance over conventional approach.