Colab notebook: https://colab.research.google.com/drive/16T5avz3zOUNcIbF_dwfxnkZDENowx-tZ?usp=sharing
Model files: https://drive.google.com/drive/folders/1LpsUx08Z3-JgvNLPQY67y8OjE4IlP4f1?usp=sharing

This release uses Tacotron2 DDC with a combination of Universal Fullband-Melgan. Model is trained using MAI-Labs dataset subset fr_FR/by_book/female/etwa/monsieur_lecoq).

Tacotron2 model is trained for 100K steps starting from a pre-trained LJSpeech model.

Tacotron2 and the vocoder model have different sampling rates (16khz vs 24khz) and this is resolved by interpolating Tacotron2 output before feeding into the vocoder as in the sample Colab notebook above.