datasets:
- ylacombe/cml-tts
language:
- it
base_model:
- SWivid/F5-TTS
pipeline_tag: text-to-speech
license: cc-by-4.0
library_name: f5-tts
This is an Italian finetune for F5-TTS Italian only so can't speak english properly
Trained over 73+ hours of "train" split of ylacombe/cml-tts dataset with 8xRTX4090, still in progress, using gradio finetuning app using following settings:
exp_name"F5TTS_Base"
learning_rate0.00001
batch_size_per_gpu10000
batch_size_type"frame"
max_samples64
grad_accumulation_steps1
max_grad_norm1
epochs100
num_warmup_updates2000
save_per_updates600
last_per_steps300
finetunetrue
file_checkpoint_train""
tokenizer_type"char"
tokenizer_file""
mixed_precision"fp16"
logger"wandb"
bnb_optimizerfalse
Pre processing
Data extracted from the datasource has been preprocessed in its transcription. From my understanding, punctuation is important because it's used to teach to have pauses and proper intonation so it has been preserved. Original italian "text" field was even containing direct dialogue escapes (long hyphen) that has also be preserved but it contained also a hypen that was used to split a word in a new line (I don't know which process was used on original dataset to create the text transcription) and so I removed that hypens merging the two part of the word, otherwise the training was done on artifacts that didn't impacted the speech. I'm only talking about Italian data on cml-tts, I don't know if other languages are affected by this too.
Current most trained model
model_25200.safetensors (45 Epoch)
checkpoints folder
Contains the weight of the checkpoints at specific steps, the higher the number, the further it went into training. Weights in this folder can be used as starting point to continue training.