Dynamic selection of a speech enhancement method for robust speech recognition in moving motorcycle environment

摘要

We present a speech pre-processing scheme (SPPS) for robust speech recognition in the moving motorcycle environment. The SPPS is dynamically adapted during the run-time operation of the speech front-end, depending on short-time characteristics of the acoustic environment. In detail, the fast varying acoustic environment is modeled by GMM clusters based on which a selection function determines the speech enhancement method to be applied. The correspondence between input audio and speech enhancement method is learned during the training of the selection function. The SPPS was found to outperform the best performing speech enhancement method by approximately 3.3% in terms of word recognition rate (WRR).