dataset preprocess

[CreepJoye]

Checks

• This template is only for usage issues encountered.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I confirm that I am using English to submit this report in order to facilitate communication.

Environment Details

python=3.10

Steps to Reproduce

Hi there, thanks for your amazing work on open-sourcing F5-TTS! While training F5 on my own dataset, I noticed an issue: the generated audio quality drops significantly when the dataset contains some audio clips longer than 20 seconds.

I saw there's a slice function in src/f5_tts/train/finetune_gradio.py that can split audio based on silent segments. So, my question is: when working with longer audio clips (e.g., over 20 seconds), do I need to slice them into shorter segments, translate them, and then use them for training? Or is it okay to include a small number of long audio clips directly? I'm planning to use a large dataset later, and slicing and translating all long audio files might be too much work.

✔️ Expected Behavior

No response

❌ Actual Behavior

No response

[SWivid]

Hi @CreepJoye ,
it's ok to do with audios of <30 total length, mainly for #719