FireRedAudio 参考音频 ASR 逐字稿 · T8star-Aix

Zero-shot voice cloning is dramatically better when the model knows what the reference clip says. FireRedAudio can auto-transcribe a missing reference transcript itself, but this node makes that step explicit and visible: feed it a reference clip, get back the same clip plus a word-perfect transcript, and wire that transcript into your TTS or voice profile. The pack's own docs call it the "one-key" reference ASR node, and it exists so the transcript is a real, inspectable part of your graph instead of a silent magic step.

How it works

It's a single asr call in the isolated worker: the audio gets converted to WAV and sent with a plain "Transcribe speech to text." prompt. max_new_tokens (default 512) caps the output length - raise it for long references, though a clone reference should rarely need much. If the worker returns an empty transcript the node raises rather than passing garbage downstream, which is the right failure mode: an empty transcript feeding a clone is worse than an error.

Outputs:

• 参考音频 - your clip, passed through byte-for-byte. The node never touches the source.
• 自动逐字稿 - the transcript string. This is the one you wire onward.
• ASR 报告 - JSON with the worker details and timing.

Where it fits - and the smart move

Two wiring patterns. The boring one: reference clip → this node → transcript into prompt_text on the TTS/SeedAudition node, audio into prompt_audio. That alone is worth doing, because a correct reference transcript is one of the cheapest quality wins in the whole TTS stack - it's the same effect the FireRedTTS3 community found: prompt_text materially improves cloning fidelity.

The smarter pattern: connect the transcript output to a VoiceProfile (the pack's reusable voice-timbre archive). A VoiceProfile that carries its own verified transcript means every future TTS run that uses that profile gets the transcript for free, and the 参考音频 ASR 逐字稿 node is the intended way to seed it. This node's real job is closing that loop: verify a candidate from the review board actually says what you think it says, then freeze the transcript into the profile. It also pairs with the ReferenceCandidates flow - after the ASR proxy ranks candidates, this node produces the authoritative transcript you trust.

Installing and the one caveat

Standard pack install: ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py, with models in ComfyUI/models/TTS/FireRedAudio/. Needs the runtime connected, and lite profile suffices (ASR is main-model-only). The caveat is expectations: the transcript is the model's best read of the clip, not a ground-truth proof. For clean studio speech in zh/en it's typically very good, but a noisy or heavily accented reference can produce a transcript that "mostly" matches - and a mostly-wrong prompt_text is worse than a blank one, because the model will pattern-match on the wrong words. If the reference is messy, read the transcript before trusting it with a clone.