Generate subtitles

The batch endpoint emits subtitle formats directly — no timestamp math, no post-processing:

curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
  -F model=openai/whisper-1 \
  -F response_format=srt \
  -F file=@episode.mp3 > episode.srt

Swap srt for vtt for the web-native format:

<video controls src="/episode.mp4">
  <track default kind="captions" src="/episode.vtt" srclang="en" />
</video>

Model choice

Any batch model works; the difference is timing granularity:

• Word-timestamp models (openai/whisper-1, assemblyai/universal-2, speechmatics/*, elevenlabs/scribe_v2, groq/whisper-large-v3) cue subtitles at word precision.
• Models without word timing degrade to one cue per utterance — fine for short clips, rough for long-form.

Speaker-labeled captions

Want [SPEAKER 1] prefixes? Request verbose_json with diarization=true and template your own cues from words[] — speakers arrive as integers:

curl -s … -F model=assemblyai/universal-2 -F diarization=true \
  -F response_format=verbose_json -F file=@panel.mp3 \
  | jq -r '.words[] | "\(.speaker)\t\(.start)\t\(.word)"'

Long files

Up to 250 MB per request. For feature-length audio, mistral/voxtral-mini accepts up to 3 hours in one shot at $0.003/min.