Repetitive output issue during long-form song ASR

[LucaLiu1313]

hi team, first of all, great work on this project! I'm testing the model's ASR performance on full-length songs. However, I encountered a severe repetitive looping issue where the model consistently outputs the same phrase. Notably, this behavior persists regardless of whether sampling is enabled or disabled, could you please provide some guidance on how to resolve this

[wx9Songs]

Previously, using Hugging Face for inference might have caused some issues, leading to repetitive looping in outputs. The model works fine with SGLang. However, we’ve just fixed the Hugging Face inference, and it should work properly now. You can refer to the infer.py script for guidance on running it.

Thank you very much for using the model and for your valuable feedback!