The new later checkpoint ruined a few of the voices I created.
The new later checkpoint mentioned here: https://huggingface.co/coqui/XTTS-v2/commit/bb2db88a1ce905424cb01fe08b1b7f6c654c6b0d
Has made a lot of my previously good voices sound really bad, and now they say some extra gibberish too.
Agree. The same CLI command with the same text and the same speaker. The result is now no more natural.
The test command :
tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 --text "Trois personnes ont été arrêtées mercredi à Edimbourg, soupçonnées d’avoir tenté d’endommager la pierre du destin pour protester contre la pauvreté, un bloc de grès utilisé depuis des siècles pour le couronnement des monarques britanniques." --speaker_wav /home/Voices/Voice01.mp3 --language_idx fr --use_cuda true
Arghh, can't upload the mp3 files, sorry
This is already discussed in many places. I recommend checking the older discussions.
Anyone find a link to any of the older discussions? A lot of threads are linking to this discussion but there's nowhere else to go from here.
I didn't find any of these discussions and no explanation. I don't know why they downgraded their model. So strange, v1.1 of the model is now better than 2.x