V2.5 Model : Fine tune of my V2 model on all CommonVoice dataset (517k sample) on 2.5k step (batch size 200), Voice cloning has improved a bit but is still not great. However, if you fine tune this model on your own personality dataset then you can get pretty good results. A good V3 model would be to fine tune for like 50k steps on this dataset and I think there would be a way to get good results but I won't try
V2 Model :
Tortoise base model Fine tuned on a custom multispeaker French dataset of 120k samples (SIWIS + Common Voice subset + M-AILABS) on 10k step with a RTX 3090 (~= 21 hours of training), with Text LR Weight at 1 Result : The model can speak French much better without an English accent but the voice clone hardly works
V1 Model :
Tortoise base model Fine tuned on a custom multispeaker French dataset of 24k samples (SIWIS + Common Voice subset) on 8850 step with a RTX 3090 (~= 19 hours of training)
Inference :
- You can use the model by downloading the "V2_9750_gpt.pth" model and use it in the tortoise-tts optimized forks (git.ecker.tech/mrq/ai-voice-cloning | 152334H/tortoise-tts-fast)
Fine tuning :
- I used 152334H/DL-Art-School for training, if you want to resume training from my epoch, follow its documentation and download "V2_9750.state"