Docs
Transcription
Two ways in: sync for short files (25 MB or less) and async batch for everything else.
Distributed processing. Jobs run on a network of vetted consumer GPUs and Macs owned by independent contributors. Do not send personal or sensitive data. Read the data policy.
Sync: POST /v1/audio/transcriptions
Multipart form, same as OpenAI.
| Field | Notes |
|---|---|
file | Up to 25 MB. Larger files return 413 file_too_large with a hint to use jobs. |
model | whisper-large-v3-turbo, whisper-large-v3 or whisper-small |
language | Optional ISO-639-1 code. Auto-detected if omitted. |
response_format | json (default), text, srt, verbose_json, vtt |
temperature | Only 0 is accepted (results must be reproducible enough to verify) |
timestamp_granularities[] | segment, word |
Async: POST /v1/audio/transcriptions/jobs
Upload a file (up to 2 GB) or pass a public https:// URL. Up to 4 hours of audio per file. See Batch API.
How audio is processed
- We fetch or receive your file and convert it on our servers to 16 kHz mono FLAC. Devices never decode your original container.
- The audio is cut into chunks of 5–10 minutes with a 1-second overlap.
- Chunks run on eligible devices, some are re-checked, and the transcript is stitched back together.
- You're billed for the duration of the converted file, measured on our servers.
Diarization
whisper-large-v3 includes speaker labels in verbose_json output. It runs only on GPUs with 8 GB of VRAM or more.