Hands-free speech recognition using a reverberation model in the feature domain

A novel approach for robust hands-free speech recognition in highly reverberant environments is proposed. Unlike conventional HMM-based concepts, it implicitly accounts for the statistical dependence of successive feature vectors due to the reverberation. This property is attained by a combined acoustic model consisting of a conventional HMM, modeling the clean speech, and a reverberation model. Since the HMM is independent of the acoustic environment, it needs to be trained only once using the usual Baum-Welch re-estimation procedure. The training of the reverberation model is based on a set of room impulse responses for the corresponding acoustic environment and involves only a negligible computational effort. Thus, the recognizer can be adapted to new environments with moderate effort. In a simulation of an isolated digit recognition task in a highly reverberant room, the proposed method achieves a 60% reduction of the word error rate compared to a conventional HMM trained on reverberant speech, at the cost of an increased decoding complexity.