The Effect of Reverberation on Optimal GMM Order and CMS Performance in Speaker Verification Systems
Noam R. Shabtai, Boaz Rafaely, Yaniv Zigel · Sciyo eBooks · 2010
The effect of GMM order on SVR with reverberant speech was investigated. Time-smearing of the feature vectors due to reverberation reduces the optimal GMM order in terms of minimum BIC and KIC. When tested on a GMM-based SVR system, reducing model order improves system performance for highly reverberant speech. A future adjustment to AGMM may be proposed in this direction The effect of room volume and RT on the performance of CMS applied to MFCC feature vectors in SVR was investigated. It was shown that the performance of CMS may degrade with the increase of RT. In some cases of high RT, CMS may increase the EER of SVR rather than decrease it. Hence, in these cases, CMS should not automatically be used. As a future work, we purpose combining a CMS decision block in SVR.