Gaussian Mixture Model Weight Supervector Decomposition and Adaptation

This report proposes a novel approach for Gaussian Mixture Model (GMM) weights decomposition and adaptation. This modeling suggests a new low-dimensional utterance representation method, which uses a simple factor analysis similar to that of the i-vector framework. The suggested approach is applied to the Robust Automatic Transcription of Speech (RATS) language identification evaluation corpus, where the speech recordings are from highly degraded communication channels. In our experiments, after modeling each utterance using the proposed approach, a Deep Belief Networks (DBN) is utilized to recognize the language of utterances. The assessment results show that the proposed method improves conventional maximum likelihood weight adaptation. It is also shown that the absolute and relative improvement obtained by the score-level fusion of the i-vector framework and the proposed method are 5% and 17% respectively.