REAL-TIME STT
Live audio to text, a chunk at a time.
Send each 5 to 10 second chunk to Whisper Large v3 Turbo as it is recorded and get timestamped text back, billed per second of audio, not per request.
Hand this to your agent
Add a Runcrate model to this project: run curl -fsSL"https://runcrate.ai/agents.md?intent=build"and follow it.
Your agent asks for an API key. Create a key
Quick start
Transcribe a stream as it arrives. Or read the request it sends.
Your agent writes and runs it. The same key works for every model and every GPU.
Available models
Models you can call today. Billed prices, as charged.
| Model | Maker | Price | Detail |
|---|---|---|---|
| openai/whisper-large-v3-turbo | OpenAI | $0.00021 per minute | Whisper's faster decoder, for live chunks |
| openai/whisper-large-v3 | OpenAI | $0.000473 per minute | The full model, when accuracy matters more than speed |
| mistralai/Voxtral-Small-24B-2507 | Mistral AI | $0.00315 per minute | Mistral's speech model, multilingual |
Why Runcrate
Built for production. One key, one balance.
Billed per second of audio
A 5-second chunk is charged 5 seconds at the per-minute price, with no per-request minimum, so a stream cut into short chunks costs what one long file would.
Context across the cut
Send the previous chunk's text in the `prompt` field. Whisper reads it as what came before, so names and spellings carry over the boundary.
Captions straight out
`response_format` srt or vtt returns caption files with timestamps; verbose_json returns segments with a start and end for each.
Skip language detection
Pass `language`, for example en, and every chunk is read as that language. A code Whisper does not know is dropped, never an error.
FAQ