A Model-based Approach to Speaker Identification using Class-specific Dictionaries
Imran Naseem, Roberto B. Togneri, Mohammed Bennamoun · UWA Profiles and Research Repository (University of Western Australia) · 2012
In this research we propose a novel speaker identification algorithm by formulating the pattern recognition task as a problem of linear regression. Essentially the concept of GMM mean supervector is used to transform variable-length utterances to fixed-length feature vectors. Training utterances from each speaker are used to develop class-specific dictionaries. The unknown test utterance is linearly modeled with each subspace and decision is ruled in favor of the speaker model with the minimum reconstruction error. Experiments on the TIMIT [1] database has shown efficacy of the proposed approach compared to the state-of-art approaches. Index Terms: nearest subspace classification, speaker identification, linear regression