Speaker Vector-Based Speaker Recognition with Phonetic Modeling

This chapter describes anchor model-based speaker recognition with phonetic modeling. Gaussian Mixture Models (GMMs) have been successfully applied to characterize speakers in speaker identification and verification when a large amount of enrolment data to build acoustic models of target speakers is available. However, a small amount of enrolment data of as short as 5 sec. might be preferred for some tasks. A conventional GMM-based system does not perform well if the amount of enrolment data is limited. In general, 1-minute or more of enrolment data are required in the conventional system. In order to solve this problem, a speaker characterization method based on anchor models has been proposed. The first application of the method was proposed for speaker indexing (Sturim et al., 2001). And the method has been also used for speaker identification (Mami & Charlet, 2003) and speaker verification (Collet et al., 2005). In the anchor model-based system, the location of each speaker is represented by a speaker vector. The speaker vector consists of the set of the likelihood between a target utterance and the anchor models. It can be considered as a projection of the target utterance in a speaker space. One of the merits of this approach is that it is not necessary to train a model for a new target speaker, because the set of anchor models does not include the model of target speaker. It can save users time to utter iteratively for model training. However, there is a significant disadvantage in the system because the recognition performance is insufficient. It has been reported that an identification rate of 76.6% was obtained on a 50-speaker identification task with 16-mixture GMMs as anchor models (Mami & Charlet, 2003). Also, an equal error rate (EER) of 11.3% has been reported on speaker verification task with 256-mixture GMMs (Collet et al., 2005). Compared with the conventional GMM approach, the performance of anchor model-based system is remarkably insufficient. The aim of this work is to improve the performance of the method by using phonetic modeling instead of the GMM scheme as anchor models and to develop text-independent speaker recognition system that can perform accurately with very short reference speech. A GMM-based acoustic model covers all phonetic events for each speaker. It can represent an overall difference in acoustic features between speakers, however, it cannot represent a difference in pronunciation. Consequently, we propose the method to detect the detailed difference in phonetic features and try to use it as information for speaker recognition. In order to detect the phonetic features, a set of speaker-dependent phonetic HMMs is used as O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg