Discriminative training of Gaussian mixture speaker models: A new approach

Abstract

Conventional speaker recognition systems use Gaussian mixture models (GMM) to model a speaker's voice based on the speaker's acoustic characteristics. This method is categorized as a non-discriminative training process, as the model-building process does not take into account the negative examples of the speaker. To increase the discriminative properties of a GMM for each speaker, a new approach that includes both positive and negative examples during the speaker training process is proposed. In this approach, speaker models are trained by moving the mixture model's means in such a way as to maximize the likelihood of speaker data while also minimizing the likelihood of negative examples for the speaker. The effectiveness of this approach on classification accuracies on speaker recognition tasks is tested on the NTIMIT database and NIST SRE 2003 corpora. The results indicate improvements in the performance of the system built using this new approach when compared to the traditional GMM-based speaker recognition systems.