8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Automatic Adaptation of the MoMel F0 Stylisation Algorithm to New Corpora

Salma Mouline (1), Olivier Boeffard (2), Paul Bagshaw (1)

(1) France Telecom R&D, France
(2) Université de Rennes 1, France

The paper investigates the adaptability of the MoMel (Modelling of Melody) Algorithm, Hirst & Espesser 1983, to new corpora. A detailed overview of the MoMel algorithm and its parameters are presented. The generality of the default parameter values to new corpora is studied empirically. Two of the parameters, related to window durations, are discovered to be highly corpus dependant. The paper presents a significant reduction in the modelling error (r.m.s. from 8.45 to 6.19 Hz) by automatically adapting these two parameters to the new corpora. Despite this optimisation, the observed errors remain perceptually noticeable (JND of approximately 3-4 Hz) and future work needs to address this problem.

Full Paper

Bibliographic reference.  Mouline, Salma / Boeffard, Olivier / Bagshaw, Paul (2004): "Automatic adaptation of the momel F0 stylisation algorithm to new corpora", In INTERSPEECH-2004, 753-756.