Batch transcription

POST https://api.speechrouter.ai/v1/audio/transcriptions — multipart form, OpenAI-compatible shape.

Form fields

Field	Default	Meaning
model	(required)	Model slug with batch mode, e.g. openai/whisper-1
file	—	The audio (up to 250 MB). Exactly one of file / url
url	—	Public URL the gateway fetches instead
response_format	json	json · verbose_json · srt · vtt · text
language	auto	Language hint
diarization	false	Speaker labels (model support varies)
keyterms	—	Comma-separated bias terms
provider_params	—	JSON object string, passed through to the vendor
include_raw	false	Include the raw provider payload (verbose_json)

Output formats

• json — { "text": "…" }, nothing else. The pipe-into-jq format.
• verbose_json — text plus words[] (timings, confidence, speakers), duration, language, and provider_raw when requested.
• srt / vtt — ready-to-serve subtitle files.
• text — the bare transcript.

curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
  -F model=speechmatics/melia-1 \
  -F response_format=verbose_json \
  -F file=@interview.mp3 | jq '.words[:3]'

Speaker labels

With diarization=true, models that return word-level speakers give you words[].speaker as integers. Models that diarize at segment level (e.g. openai/gpt-4o-transcribe-diarize) return one word entry per speaker turn.

Timeouts & sizes

Uploads and remote URLs are capped at 250 MB. Batch requests hold the connection until the provider finishes — allow up to 10 minutes for very long audio in your HTTP client timeout.