SenseVoiceSmall adapted to anime Japanese (encoder + CTC head)

Fine-tuned from FunAudioLLM/SenseVoiceSmall with CTC only and MUSAN music mixed into the training audio, 8000 steps, on 92.28 h of Japanese anime clips from ANIM-400K (clips with pretrained-CTC CER > 0.8 against the annotation were dropped, 14596 of 66722 clips).

model.pt = best mean validation CER over raw and music-mixed conditions (step 8000), model_final.pt = last step. Load like the original model, e.g. AutoModel(model=<local folder>).

Listen

Four random clips from the test shows (never seen in training; fixed seed, not cherry-picked). Greedy CTC through FunASR, Japanese, no ITN. CER is against the reference with punctuation ignored (lower is better).