This model was trained for 4.2 million steps on the german Mozilla common voice 19.0 recordings and an internal dataset. It is designed for text-to-speech synthesis in German.
The command to train the model is:
accelerate launch --mixed_precision=bf16 finetune_cli.py --exp_name F5TTS_Base --learning_rate 1.8e-05 --batch_size_per_gpu 8000 --batch_size_type frame --max_samples 0 --grad_accumulation_steps 1 --max_grad_norm 1 --epochs 40 --num_warmup_updates 2000 --save_per_updates 100000 --last_per_steps 10000 --dataset_name german_speak --finetune --pretrain ckpts/german_speak/model_last.pt --tokenizer pinyin --log_samples --logger wandb
The checkpoint supports German and can be
downloaded here.
