🎙️ Wuxue Speech Recognition

Original Fun-ASR-Nano versus our experimental 36-clip, 72-step adapter. Record or upload up to 15 seconds. Free ZeroGPU may have cold starts, queues, and daily usage limits.

Privacy: audio and optional reference text are processed on Hugging Face. No intentional dataset saving or transcript logging. Temporary working audio is deleted after inference; upload cache is periodically cleaned after 60 seconds.

Record or upload audio

Optional reference transcript — scoring only

Original Nano

Our adapter — 36 clips / 72 steps

Score and status

Recognition may be inaccurate. The 36-clip adapter is experimental: its earlier reused exploratory evaluation CER was 63.33% versus 61.67% for the base, so improvement has not been demonstrated. Both outputs use the same audio, dither-zero frontend and greedy decoding. The four-clip memorization adapter is not used. Reference text never enters the model prompt.