Text independent speaker identification using gaussian mixture model

Abstract

This paper describes text-independent (TI) Speaker Identification (ID) using Gaussian mixture models (GMM). The use of GMM approach is motivated by that the individual Gaussian components of a GMM are shown to represent some general speaker-dependent spectral shapes that are effective for speaker identity modeling. For speaker model training, a fast re-estimation algorithm based on highest likelihood mixture clustering is introduced. In this work, the GMM is evaluated on TI Speaker ID task via series of experiments (model convergence, effect of feature set, number of Gaussian components, and training utterance length on identification rate). The database consisted of Malay clean sentence speech database uttered by 10 speakers (3 female and 7 male). Each speaker provides the same 40 sentences utterances (average length- 3.5s) with different text. The sentences for testing were different from those for training. The GMM achieved 98.4% identification rate using 5 training sentences. The model training based on highest likelihood clustering is shown to perform comparably to conventional expectation-maximization training but consumes much shorter computational time.