search

FunAudioLLM / CosyVoice

Multi-lingual large voice generation model, providing inference, training and deployment full-stack ability.
https://funaudiollm.github.io/
Apache License 2.0
6.1k stars 654 forks source link

Try to train llm model with cosyvoice.fromscratch.yaml. The loss is not converging. #391

Open cx16528 opened 1 month ago

cx16528 commented 1 month ago

Try to train llm model with cosyvoice.fromscratch.yaml. The loss is not converging. There are the 180th epoch loss and acc. Snipaste_2024-09-13_10-33-16 The loss still more than 3.0 and the acc cannot increase more than 0.30.

aluminumbox commented 1 month ago

acc 0.25 is ok on libritts, check if you can inference with it

KunZhou9646 commented 4 weeks ago

Same as my situation. However, the synthesized speech is not of good quality.