Finetune Guide

[tamerkanak]

Can you add fine tune code/guide for OpenMOSS-Team/MOSS-TTS-Realtime, OpenMOSS-Team/MOSS-TTS?

[YWMditto]

We're building this, please stay tuned!

[YWMditto]

We have already released an open-source tutorial on fine-tuning the MossTTSDelay architecture, and will subsequently provide corresponding versions of MossTTSLocal and MossTTSRealtime.

https://github.com/OpenMOSS/MOSS-TTS#fine-tuning

[omarabb315]

thank you for the tutorial, what is the recommended length distribution for the audio segment in training phase?,
I tried 2-15s, 2-30s, 2-200s but every time I had the early clipping problem

[tekinek]

thank you for the tutorial, what is the recommended length distribution for the audio segment in training phase?, I tried 2-15s, 2-30s, 2-200s but every time I had the early clipping problem

@omarabb315 With my single speaker dataset (1~30s, 100 hours), the finetuned moss-tts-delay could speak clearly after an epoch, but early clippping happens on almost every test sentence. Please share your experience if you have solved the issue.

[omarabb315]

I am still waiting for the solution actually, I tried increasing the segments range to be (2-200s) improved a bit, but still suffering from the same issue ...

[tekinek]

I am still waiting for the solution actually, I tried increasing the segments range to be (2-200s) improved a bit, but still suffering from the same issue ...

During inference, manual setting of text_temperature and text_temperature in model.generate() helped.

model.generate(text_temperature=0.8, text_top_p=0.95, ...)