MOSS-TTS Whisper Transcribe

Whisper Transcribe is the least glamorous node in the MOSS-TTS pack, and the one that quietly unblocks the most interesting workflow it has. Its entire job here is to turn audio into text - a STRING transcript - so you don't have to type the transcript yourself when using Continue Speech. Feed it the prefix clip, wire its transcript output into Continue Speech's prefix_text, and you've automated the fiddly part.

The name is not a gimmick: this is genuinely local OpenAI Whisper via Hugging Face Transformers, same Whisper everyone's been running for years. Standalone, it's a perfectly usable ComfyUI Whisper node for transcription or translation of anything - it just also happens to be wired into the MOSS workflow.

How it works

It converts the ComfyUI AUDIO to a numpy array and runs it through a Transformers Whisper pipeline. Models live in ComfyUI/models/audio_encoders/ - the same folder other audio packs share - and get downloaded on demand into the catalog. It handles chunking for longer clips and has a small piece of sanity handling: if the input audio is silent or empty, it returns an empty string instead of hallucinating a transcript.

The inputs that matter

• audio - whatever you need transcribed, as ComfyUI AUDIO.
• model - five choices: whisper-large-v3-turbo is the default and is fast plus usually accurate enough for prefix transcripts. whisper-large-v3 is heavier and better on accented or noisy reference audio; medium, small, tiny exist if you want speed over accuracy.
• dtype - auto picks bf16 on supported CUDA/XPU and fp32 otherwise; the manual options are bf16 and fp32.
• language - auto detects it, but setting it explicitly improves accuracy, and for the MOSS use case the transcript needs to be right.
• task - transcribe keeps the original language; translate outputs English.
• chunk_length_s - 30 by default; 0 lets Transformers decide. Matters for longer clips.
• download_if_missing - true by default; downloads the chosen Whisper model into audio_encoders when it isn't there.

Output: transcript - a plain STRING with the recognized text, meant for Continue Speech's prefix_text but usable anywhere a string goes.

Installing

Pack-wide install, same as the others: ComfyUI Manager → search MOSS-TTS, or

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/Moss_TTS-ComfyUI.git
../venv/bin/python Moss_TTS-ComfyUI/install.py

then restart. The Whisper model download happens on first run of the node, not at install time.

Common issues

Remember what this node produces. It's a raw ASR transcript, so punctuation is unreliable - and Continue Speech concatenates prefix_text and continuation_text directly, which is exactly where a missing period turns into a run-on sentence. A quick edit between the Whisper output and the continuation fixes the seams.

For most prefix audio, whisper-large-v3-turbo is the right default - it's what the pack author ships, and the transcripts only feed a continuation alignment, not a public-facing caption. Reach for whisper-large-v3 when the reference is accented, quiet, or noisy. If the transcript comes back empty, the input was silent or near-silent - check the audio, not the node. And if you're running low on VRAM, note that Whisper is a separate model loading alongside the MOSS stack, so a smaller Whisper model (or fp32→bf16) is a legitimate lever.