In some previous works on the speaker identification , we have underlined and quantified the influence of the context (enunciation situation, coarticulation) one someone's phonetic realisations. These results - recognition of 1 out of 20 speakers, with a safety margin over 99%, using from 5 to 7 phonemes - led us to check the reliability of such a process within the context of automatic speaker recogntion. Therefore, we have created a sound data base (each of the sixty speakers pronounced several hundreds sentences) during many mounth-lasting recordings. Selected speakers were not trained to do this kind of exercise and had very similar socio-cultural and age characteristics. There were usual recording condition (offices, far end corridor, etc.). After an speaker independant automatic localisation of phonemes, we drew up statistics on intra and inter-speaker variability using several spectral shape algorithms. Afterward we have measured tfie discrepencies due to the inaccuracies of the phoneme automatic localisation. The outcome shows a drop in the distances selective ability in comparison with the previous tests. A sharp fall mainly due to the phonemes automatic localisation's randomnessand DB recording type's uncertainty. Such a fact does not prevent the use of these techniques on a speaker recognition system given the passable performances.
Bibliographic reference. Bonastre, Jean-Francois / Meloni, Henri (1993): "Automatic speaker recognition and analytic process", In EUROSPEECH'93, 441-444.