First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Extraction of Phoneme-Dependent Individuality Using HMM-Based Segmentation for Text-Independent Speaker Recognition

Hideki Noda, Masuzo Yanagida

Kansai Advanced Research Center, Communications Research Laboratory, Kobe, Japan

This paper describes a new method for text-independent speaker recognition which exploits phoneme-dependent voice individuality without direct phoneme recognition. This method uses a segmentation technique based on the Hidden Markov Model (HMM). Appropriate segmentations are expected to be carried out through parameter estimation of models, given enough amount of utterances with their phonetic transcriptions. Segmentations being completed, the difference between feature vectors of reference and input which belong to the same phoneme is used as the dissimilarity measure for speaker recognition. Speaker verification performance has been evaluated by experiments using 20 word utterances of 177 male speakers. In a experiment 95.4% verification rate is achieved using the proposed method, whereas 89.3% by a well-known VQ method.

Full Paper

Bibliographic reference.  Noda, Hideki / Yanagida, Masuzo (1990): "Extraction of phoneme-dependent individuality using HMM-based segmentation for text-independent speaker recognition", In ICSLP-1990, 129-132.