IMPROVING MICROPHONE ARRAY SPEECH RECOGNITION WITH COCHLEAR IMPLANT-LIKE SPECTRALLY REDUCED SPEECH

Cochlear implant-like spectrally reduced speech (SRS) has previously been shown to afford robustness to additive noise. In this paper, it is evaluated in the context of microphone array based automatic speech recognition (ASR). It is compared to and combined with post-filter and cepstral normalisation techniques. When there is no overlapping speech, the combination of cepstral normalization and the SRS-based ASR framework gives a performance comparable with the best obtained with a non-SRS baseline system, using maximum a posteriori (MAP) adaptation, either on microphone array signal or lapel microphone signal. When there is overlapping speech from competing speakers, the same combination gives significantly better word error rates compared to the best ones obtained with the previously published baseline system. Experiments are performed with the MONC database and HTK toolkit.