Resource-constrained multilingual speech translation

Starting: 01/11/2022
PhD Student: Hugo Riguidel
Advisor(s): Antoine Laurent (LIUM)
Co-advisor(s): Anthony Larcher (LIUM), Josep Crego (Systran)
Funding: CIFRE - ministère de l'Enseignement supérieur, de la Recherche et de l'Innovation


This thesis work is in the field of automatic language processing and more specifically in automatic speech translation. The usefulness of machine translation for the community is no longer in question: allowing citizens to dialogue while speaking in their native language is a way to facilitate interactions between humans.


One of the goals of the thesis is to be able to encode both semantic and phonetic information, in a multi-lingual way, in order to perform the speech translation task. The idea behind the use of a single model is to have a robust model that can work in a scenario where few annotated resources exist for a given language.
The work will be done in collaboration with a research engineer, one of whose objectives will be to integrate the work into a video-conference type demonstrator. In its final version, the model should be able to be used live (streaming). The size of the model will therefore be of interest.