Text-to-speech and voice conversion are two distinct speech generation techniques. Text-to-speech  (TTS) is a process that generates speech from a sequence of graphemes or phonemes. Voice conversion is the conversion of speech from a source voice to a target voice. These processes find their application in domains such as Computer Assisted Language Learning, for example.

 

However, these two processes share some bricks, particularly the vocoder that generates speech from acoustic characteristics or the spectrogram. The quality of these two technologies has been significantly improved thanks to the availability of massive databases, the power of computing machines, and the deep learning paradigm implementation. On the other hand, restoring or controlling expressiveness, and more generally considering suprasegmental information, remains a major challenge for these two technologies.

 

This internship topic aims at setting up a common framework for both technologies. We aim at a joint deep learning framework to generate speech (target voice) from either speech (source voice) or text.

 

It will be supervised by members of the EXPRESSION team (IRISA): Aghilas Sini, Pierre Alain, Damien Lolive, and Arnaud Delhay-Lorrain.

 

Please send your application (CV + cover letter) before 10/01/2022 to  à aghilas.sini@irisa.fr,palain@irisa.fr, damien.lolive@irisa.fr, arnaud.delhay@irisa.fr

Start date: 01/02/2022 (flexible)

======================