Load Whisper (mtb)

This is the loader that puts OpenAI's Whisper - the go-to open speech-recognition model - into your ComfyUI graph. On its own it does nothing visible; it loads a model and hands it downstream. It's step one of a two-node dance: Load Whisper here, then Audio To Text to actually transcribe. Keeping the load separate means the model stays in memory across runs instead of reloading every time you transcribe, which matters once you're iterating.

Why transcribe inside ComfyUI at all? Subtitles and captions, mostly. Pulling a text transcript off a voice track so you can burn subtitles onto a video, feed the words into a downstream text node, or time animation to speech. As audio-driven video work has grown, having Whisper right in the graph beats round-tripping through a separate tool.

How it works

It downloads (or loads from cache) a Whisper checkpoint and its processor, wraps them into a pipeline object, and outputs that. Whisper itself is an encoder-decoder trained on a huge pile of multilingual audio, which is why it's robust to accents, background noise, and code-switching in a way older ASR wasn't.

The inputs that matter

• model_size - the accuracy-versus-cost dial, and the only choice you really make. The ladder runs tiny, base, small, medium, large, large-v2, large-v3, large-v3-turbo, plus an English-only medium.en. Bigger = more accurate, more VRAM, slower. tiny is fine for clean English and downloads in seconds (~75MB); the large family is a few GB and much better on hard audio. large-v3-turbo is the sweet spot for most people - nearly large-v3 quality at a fraction of the speed cost. If you only ever do English, medium.en punches above its size.
• download_missing (default false) - and this default is a trap. If the model isn't already in your cache and this is off, the node errors instead of fetching it. Tick it on for the first run so it downloads what it needs.

The single output is pipeline (WHISPER_PIPELINE). It goes straight into the Audio To Text node.

How to install it

ComfyUI Manager: search MTB Nodes (comfy_mtb), install, restart. Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/melMass/comfy_mtb and restart. The Whisper nodes need the underlying Whisper/transformers stack installed, which is an optional dependency for this pack. MTB loads gracefully - if a node's dependency is missing, that node quietly fails to load while the rest work, and it points you at http://127.0.0.1:8188/mtb, a dependency panel that installs the missing pieces. If Load Whisper doesn't show up in your node list, that's where you go.

Common issues

The number-one confusion is download_missing being off by default and the node erroring on a model you haven't fetched yet. Turn it on, run once, and the model lands in cache; you can leave it on with no downside.

The number-two issue is picking a model your GPU can't hold. large-v3 on a small card will spill or OOM - drop to large-v3-turbo or small. And remember this node just loads the model; if transcription quality is bad, that's a model-size or language question for the Audio To Text node downstream, not something to fix here.