ThinkingLLM Whisper ASR

ComfyUI has plenty of ways to make noise; it has few ways to listen. The Whisper ASR node is the transcription half of the ThinkingLLM pack's audio story - speech-to-text via faster-whisper that turns a connected AUDIO input or a file path into a TRANSCRIPT string you can pipe into any text node in the graph. It's the missing front end for every local-LLM workflow in this pack: Whisper transcribes, a Qwen node summarizes.

How it works

You feed it either an audio connection (from a Load Audio node) or an audio_file_path pointing at M4A, MP3, WAV, FLAC, or anything FFmpeg can read. The node decodes it to 16 kHz mono WAV, runs faster-whisper, and returns three strings:

• TRANSCRIPT - the clean, joined text.
• SEGMENTS_JSON - the per-segment timings and text, which is what you want if you're building subtitles or aligning transcript to video.
• RAW_TRACE - the diagnostic trail: source, model, device, compute type, detected language and confidence, duration.

The defaults are tuned for Windows reliability, and that's the right call: small model on cpu with int8 compute. It's not fast, but it works on a machine with no CUDA setup at all. Switch device to cuda (and compute_type to float16 or int8_float16) once your CTranslate2 CUDA runtime is confirmed working.

The inputs that matter

• model_size - tiny / base / small / medium / large-v3 / distil-large-v3. small is the honest starting point; large-v3 is a big quality jump and a big download; distil-large-v3 is the speed pick.
• language - auto (default) lets Whisper detect it, or pin it (en, de, fr, es, ja, ko…) to speed things up and avoid misdetection.
• task - transcribe keeps the original language; translate returns English.
• beam_size - 5 default; higher is more accurate and slower.
• vad_filter - on by default; skips long silences and non-speech before transcribing. Leave it on unless it's eating the first words of a quiet recording.

Installing

Manager (search ThinkingLLM) or:

cd ComfyUI/custom_nodes
git clone https://github.com/goodguy1963/ComfyUI-ThinkingLLM.git
cd ComfyUI-ThinkingLLM
pip install -r requirements.txt

The pack's requirements declare faster-whisper, and the node returns an install hint in its output if the library is missing. Two real dependencies: FFmpeg must be available for audio decoding, and the first run downloads the selected Whisper model into the Hugging Face cache - on a dial-up mood that's a 150 MB–3 GB surprise for the bigger models.

Gotchas

If you get an install-hint message out of the TRANSCRIPT output, faster-whisper isn't importable in ComfyUI's Python - that's an environment issue, not a workflow issue, and the hint names it. CPU + int8 is the reliable Windows default but genuinely slow on long files; don't benchmark the node's quality on a 30-minute recording before switching to CUDA. And remember the division of labor in this pack: Whisper gives you words, Gemma 4 Audio gives you understanding. If you only need "what was said," this is the cheap node; if you need "what's happening in this audio," that's the Gemma node's job.