Fiche de TEYTAUT Yann

Titre : On Temporal Constraints for Deep Neural Voice Alignment
Date : 7 juillet 2023
Heure: 09:30
Résumé : This thesis studies the voice alignment task, which aims to temporally synchronize a voice signal with a sequence of symbols describing its evolution. The core of this research is the development of an acoustic model, ADAGIO, capable of predicting time-symbol representations. Recent progress in deep learning have led to implement ADAGIO as a deep neural network in a powerful generic formalism: the “Connectionist Temporal Classification” (CTC). However, the great flexibility offered by CTC is undermined by its intrinsic lack of guarantees for temporally accurate predictions. Therefore, additional temporal constraints have been propoposed to improve the quality of the alignments inferred with CTC. Concretely, three ancillary tasks of (1) spectral content reconstruction; (2) audio structure propagation; and (3) guided monotony are introduced and induce a positive impact on the alignment between voices, texts, and notes. ADAGIO has contributed to many practical, collaborative applications.
Lieu : IRCAM 1 place Igor Stravinsky 75004 Paris