Instantaneous model adaptation method for reverberant speech recognition

摘要

An acoustic model adaptation algorithm is proposed for reverberant speech recognition. Inspired by the eigenvoice adaptation framework, multiple acoustic models reflecting various reverberant environments are combined for instantaneous adaptation. Using artificially generated reverberant speech, multiple acoustic models are trained according to multiple reverberation times. The mean vectors of the optimal acoustic model are obtained as a weighted sum of those of multiple acoustic models by using a maximum-likelihood criterion. For effective model combination, reverberant speech is preprocessed. Experiments on English continuous speech recognition tasks in a simulated reverberant environment show that the proposed method performs better than the conventional adaptation techniques.