7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Prosodic Parameter for Speaker Identification

Katarina Bartkova, David Le Gac, Delphine Charlet, Denis Jouvet

France TÚlÚcom R&D, France

This study investigates to what extend prosodic parameters are speaker dependent and, therefore, can be used in a speaker identification system. Fundamental frequency, phone duration and phone energy are investigated and modeled under different forms and their effi- ciency to identify successfully the speaker, to whom the model belongs, is evaluated. The speech material used to model and evaluate the prosodic parameters is collected from 28 speakers and consists of free spontaneous speech. For 61% of the speakers one or several very efficient prosodic cues are found yielding an average ranking lower, thus better, than the 3rd best position. For 36% of the speakers the average ranking of the right speaker is between the 3rd and the 5th position. Only one speaker out of 28 is not satisfactorily represented by any of the prosodic parameters.

Full Paper

Bibliographic reference.  Bartkova, Katarina / Gac, David Le / Charlet, Delphine / Jouvet, Denis (2002): "Prosodic parameter for speaker identification", In ICSLP-2002, 1197-1200.