Evaluating long-term spectral subtraction for reverberant ASR

Even a modest degree of room reverberation can greatly increase the difficulty of automatic speech recognition. We have observed large increases in speech recognition word error rates when using a far-field (3-6 feet) microphone in a conference room, in comparison with recordings from head-mounted microphones. In this paper, we describe experiments with a proposed remedy based on the subtraction of an estimate of the log spectrum from a long-term (e.g., 2 s) analysis window, followed by overlap-add resynthesis. Since the technique is essentially one of enhancement, the processed signal it generates can be used as input for complete speech recognition systems. Here we report results with both the HTK and the SRI Hub-5 recognizer. For simpler recognizer configurations and/or moderate-sized training, the improvements are huge, while moderate improvements are still observed for more complex configurations under a number of conditions.