Towards Introducing Long-term Statistics in Muse for Robust Speech Recognition

In this paper, we propose new developments of the MUltipath Stochastic Equalization techniques (MUSE). The MUSE technique is based on an enriched model of speech, composed of both a classical model of clean speech with HMM and equalization functions. This technique is able to reduce the recognition error rate due to a mismatch between the training and testing conditions. In order to track long-term variation of this mismatch, the introduction of a priori statistics on the equalization function is studied. In the case of Bias Removal, this approach has been implemented in HTK and tested on the Numbers95 database. Experiments show that the convergence of the bias computation is fast enough and limits the effect of the a priori values. However, both the fast convergence property and the proposed framework open research directions towards more complex equalization functions.