Raon Whisper Transcribe

Raon Whisper Transcribe doesn't clone any voices. It does something more boring and more important: it transcribes your reference clip so the cloning node knows what was said. Raon OpenTTS Generate needs an exact transcript of the reference audio to condition on - and the pack is upfront that this is the single biggest lever on clone quality. Typing it by hand is an option, but it's error-prone and slow, which is exactly the job this node exists for. You point it at the same clip you're using as a reference, it returns the text, and you wire that text into ref_text on the generate node. The whole workflow becomes: load audio → transcribe → clone.

How it works

It's plain Whisper ASR running locally through transformers - no cloud call, no key. It loads the selected model from ComfyUI/models/audio_encoders (downloading it there on first use, only if download_if_missing is on), transcribes your audio, and returns two outputs: the original audio passed through completely unchanged, and the transcript string.

The default model is whisper-large-v3-turbo, and it's the right default - fast enough to not slow the workflow down and accurate enough for reference transcripts. This is a case where chasing the biggest model is pointless: a transcript doesn't need to be perfect, it needs to match what the clip actually says, so the conditioning lines up.

The inputs that matter

• audio - your reference clip (AUDIO in, so it pairs with a Load Audio node).
• model - large-v3-turbo default; drop to small/tiny if you're on a weak card, though for English speech turbo is usually fine even on modest hardware.
• language - auto-detect is fine for English. Setting it explicitly (e.g. forcing english) genuinely helps accuracy, especially on clips with noise or accents - Whisper's auto-detection can trip on short clips.
• task - transcribe keeps the original language; translate forces English output. Leave it on transcribe unless you're deliberately normalizing.
• chunk_length_s - 30 s default; if your reference clip is long, either raise this or leave 0 to let Transformers pick.

The transcript output is a plain STRING, so it feeds directly into the ref_text input of Raon OpenTTS Generate. Nothing else to stitch - the node was built for exactly this wiring.

Installing it

ComfyUI Manager will find "ComfyUI-Raon-OpenTTS", or the usual:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-Raon-OpenTTS
# restart ComfyUI

It needs transformers (already in the pack's requirements) and a Whisper model, which it fetches automatically. That's the whole dependency story - this node is the lightweight end of the pack.

Where people get burned

The main trap is skipping it and hand-typing ref_text, then wondering why the clone sounds off - a mismatched transcript quietly degrades the result. The second is a quiet first-run: the Whisper model download happens inside the transcribe call, so the first run can sit there for a bit before producing anything. That's normal, not a hang. And if you're cloning English speech, don't bother switching models - turbo at auto language is the fast path and it's usually right. The one setting worth touching is language, and only when auto keeps guessing wrong.