RunPod Whisper V3 (Speech→Text)

Most of the nodes in this pack generate things. This one reads. RunPod_WhisperV3Large points at RunPod's hosted whisper-v3-large endpoint and hands you back a transcript - OpenAI's biggest open speech-to-text model, served as an API call so you don't download a single weight. If you've ever wanted a video's dialogue as text flowing into the rest of your workflow, this is the node.

Where does that actually get used? Three spots. Transcribing a clip you just rendered so you can caption or subtitle it. Feeding dialogue into a TTS or lip-sync pipeline downstream - a transcription step is how a talking-head workflow gets its script into a form another node can consume. And plain old automation: batch-transcribing audio you'd rather not listen to. It's a small, quiet utility compared to the video generators in this pack, and honestly it's the one with the least drama - the model is well-established and the failure modes are few.

How it works

Same pack machinery as every other node: it's a thin wrapper built by the pack's endpoint factory, submitting an async job to https://api.runpod.ai/v2/whisper-v3-large/run with your key in the Authorization header and polling until the job completes. The output type is text, so instead of a URL it extracts the transcript from the response and hands it to you as a STRING.

The inputs that matter

There are only three, and the first gotcha is hiding in plain sight:

• api_key - your RunPod key.
• audio - a URL string pointing at the audio file. This is the important one: there is no AUDIO tensor input on this node. You can't drag a local file in; the audio has to be reachable at a URL (hosted somewhere, or on a service that gives you a link). That's the pack's design - audio is just a string the node forwards to the endpoint.
• prompt - optional, defaults to empty. This is Whisper's own prompt parameter, used to give the model context or bias the transcription (a topic, speaker names, unusual vocabulary). Leave it blank and it just transcribes.

That's the whole list - no language field, no timestamps toggle. Whisper's language detection and everything else happen server-side with what you give it.

The output

text, a STRING, straight out. Wire it into Show Text or Save Text, or into any downstream text-consuming node. Unlike the video nodes there's nothing that expires - text is text once you have it.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

Restart, or search "RunpodNodes" in ComfyUI Manager. No extra Python dependencies and no model downloads - the pack uses only requests, PIL, torch, and numpy, all of which ComfyUI already ships. Whisper V3 Large never touches your disk; it runs on RunPod's side.

The one real annoyance is the URL requirement: if your audio isn't hosted anywhere, you'll be hunting for a file host before you can transcribe. And the standard API-node caveats apply - you're billed per job (audio runs about $0.05 per 1000 characters per the README), the audio leaves your machine, and the pack itself is a brand-new single-author release with no track record yet, so a quick skim of the source before you paste in a key is cheap insurance. For turning a URL into a transcript inside a workflow, though, it just works.