Laurent Besacier - The challenge of discovering linguistic units from raw speech

In this seminar, I will present several collective scientific projects [1,2,3] in which I was involved in 2017-2019. What do they have in common?
Discovering linguistic units from raw speech without any other supervision. Or almost…
One of these projects is the zero resource speech challenge that addresses an ambitious fundamental scientific question for AI: how can a system autonomously acquire language?
While 2015 and 2017 challenges addressed unsupervised word and subword (pseudo-phones?) discovery, the 2019 challenge proposed to build a speech synthesizer without any text or phonetic labels: hence, TTS without T (text-to-speech without text).

 

[1]https://arxiv.org/pdf/1712.04313.pdf 

[2] https://arxiv.org/pdf/1802.05092.pdf 

[3] https://arxiv.org/pdf/1904.11469.pdf