Raphael Lenain

Created a representation learning model trained on raw audio to learn de-identified representations of prosody. I was one of the co-authors of the paper, which involved an ideation phase through whiteboarding and brainstorming sessions, a development phase where we implementated the models in PyTorch using Hugging Face, and our own custom built classes. There was a refining phase during which we tested hypotheses and performed ablation studies to understand the inner workings of our approach, and finally a write up phase leading up to the conference deadline.

Designed, Implemented, and maintained an internal package which fully automated data preprocessing and Machine Learning (ML) model training and validation. Using this package made our research and deployment pipeline fully streamlined allowing fast iteration and led us to being able to experiment with ML ideas and publish papers in a very short amount of time.

Researched, designed, and implemented a novel Text-to-Speech (TTS) synthesis technique to control synthesized speech attributes (emotions, whispering, etc.). Successfully deployed the model in production, which was in use by Papercup customers, and released translated videos across various platforms (e.g., Youtube).