Isolated word recognition in reverberant environments

The additive noise and channel distortion caused by reverberation can degrade the performance of isolated word recognition(IWR), and have become the key constraint to the applications of IWR. In this paper, we present a reverberation robust isolated word recognition method. By using the relative autocorrelation sequences (RAS) based voice activity detection, influences of additive noise can be eliminated. To reduce the channel distortion Cepstral mean subtraction (CMS) is employed in Mel frequency cepstral coefficients (MFCC) extraction. And Gaussian mixture model (GMM) is used for the statistical modeling. The performance of the presented method in various reverberation conditions was evaluated by the recognition experiments.