Pre-recorded transcription — pick your engine

Two batch surfaces. They are not interchangeable.

	`POST /v1/listen/prerecorded`	`POST /v1/transcribe`
Engine	Parakeet-TDT (streaming, chunk-pumped)	faster-whisper large-v3 (batch)
Shape	Synchronous — one response	Async job — submit / poll / result
Audio length	Short clips (≲ a few minutes)	1 hour+
Languages	25 EU locales	99 incl. tr, ar + all EU
Word timestamps	Yes	Yes
Diarization	Sortformer (4-speaker cap)	Pluggable — Sortformer or pyannote
Word hints / context	Keyword boost only	`word_hints` + `context` + medical lexicon
Webhook	No	`callback_url`

Rule of thumb. Audio longer than a few minutes, or any need for Turkish/Arabic, multi-speaker labelling, or vocabulary biasing → use /v1/transcribe. The synchronous route stays for short clips and is unchanged.

Async engine status: ENABLED

Submit a job

Options ride as ONE JSON string in an options form field — not as separate -F fields. A flat -F 'language=de' is silently ignored and the job runs with defaults.

curl -X POST http://localhost:4600/v1/transcribe \
  -F 'file=@consultation.wav' \
  -F 'options={"language":"de","diarize":true,"medical_lexicon":true,"word_hints":["Ösophagusvarizen","Sankt Augustin"]}'
# → 202 {"job_id": "...", "status": "queued"}

# Remote URL: options live inline, in the SAME object as `url`
curl -X POST http://localhost:4600/v1/transcribe \
  -H 'content-type: application/json' \
  -d '{"url":"https://example.com/consultation.mp3","language":"de"}'

curl http://localhost:4600/v1/transcribe/$JOB_ID          # status + progress
curl http://localhost:4600/v1/transcribe/$JOB_ID/result   # DG Results envelope

Options

Rendered from the live request model — this table cannot drift from the code.

Field	Default	Purpose
`language`	`—`
`context`	`—`
`word_hints`	`—`
`medical_lexicon`	`False`
`smart_format`	`True`
`term_correction`	`False`
`diarize`	`False`
`callback_url`	`—`
`url`	`—`

Vocabulary biasing

context and word_hints feed Whisper's initial_prompt, which conditions the decoder toward supplied spellings. This is the highest-leverage control for medical terms, surnames and place names. medical_lexicon=true appends the in-repo lexicon for the resolved locale. Biasing is a prior, not a guarantee — it raises the likelihood of a spelling, it does not force it.

The envelope is the same Deepgram-shaped Results frame the streaming path emits — see TypeScript.