This article describes an automatic text-independent speaker recognition method based on phonetic Hidden Markov Models (HMM) [1],[2]. The challenge of our method is to reduce the duration of the test phase: only one sentence is requested for the speaker verification. We will describe how the pitch can be used to improve the automatic phoneme segmentation of a sentence and thus also the training and testing phases of the method, reducing the error rates.

We are currently studying a totally different approach that we'll briefly present: the pitch histogram method. This method is only based on the pitch values: each speaker is characterised by his/her pitch frequency histogram. It's not proposed as a complete solution by itself, but could be included as a new feature in a more complete recognition system.