Load Whisper Transcription Model

Before you can transcribe anything, you need this node. It downloads (or loads from cache) an actual Whisper checkpoint and hands the rest of the pack a ready-to-use pipeline object. Nothing visibly happens when you run it - all the interesting work happens downstream, in Whisper Transcribe. Two dropdowns, no fine-tuning knobs: this pack deliberately doesn't expose Whisper's dozen esoteric decoding parameters, just the two that actually matter to a beginner - which model, and what language.

The two choices that matter

• model_id - eleven options, and every one is a straight HuggingFace Hub identifier (openai/whisper-large-v3 down to openai/whisper-tiny.en). That tells you exactly what's running under the hood: this is HuggingFace transformers' own Whisper pipeline, downloaded and run locally on your GPU - not OpenAI's paid API. No key, no per-minute billing, but you pay in VRAM and download size instead. The ladder runs tiny → base → small → medium → large → large-v2 → large-v3, plus English-only .en variants of the smaller sizes that skip multilingual decoding and come out a bit faster and more accurate when you know the audio is English. large-v3 is the best transcription quality this pack can give you; tiny is nearly free to run and fine for testing the graph or clean, simple audio. For anything you actually care about the accuracy of, go medium or bigger if your GPU can hold it.
• language - three options: en, fr, or auto. auto is what most people want - Whisper was trained across close to a hundred languages and detects which one it's hearing on its own. en/fr force a specific decode language when you already know what you're feeding it, which is marginally faster and slightly more accurate than auto for a language you're certain about, since the model skips its own detection step.

The single output is pipeline (TRANSCRIPTION_PIPELINE) - wire it into Whisper Transcribe or Whisper Transcribe Batch, and reuse the same loaded model across as many transcribe calls as you want instead of reloading it every time.

Installing it

ComfyUI Manager: search ComfyUI-TranscriptionTools, install, restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/royceschultz/ComfyUI-TranscriptionTools

then restart. The model itself downloads from HuggingFace the first time you run this node with a given model_id - expect a real pause on that first run (tiny is under 100MB, large-v3 is a few gigabytes), and every run after is instant because it's cached.

Common issues & troubleshooting

The first run looks stuck. That's the HuggingFace download, not a crash. Check your console for progress, and if you're self-hosting somewhere locked down, confirm the executor actually has outbound internet access.

large-v3 OOMs or the graph stalls. It's a genuinely large model. Drop down the ladder - medium is a solid middle ground, small if you're on a modest card. If you know your audio is English-only, reach for the matching .en variant over the multilingual one; it's lighter for the same rough accuracy.

Import errors when the pack loads. This node pulls in transformers, and version conflicts between custom-node packages sharing one Python environment are a real, well-documented friction point in the ComfyUI ecosystem generally - everything installs into the same environment with no isolation, so the newest thing you added can quietly break an older pack's pinned version. Check the console for the actual import traceback before assuming the bug is specific to this node.