Providers and capabilities¶
Capabilities are versioned metadata, not promises of parity. list_transcription_models is the runtime source of truth for the deployed version.
| Model | URL input | Diarization | Timestamps | Styles | Hints |
|---|---|---|---|---|---|
| Microsoft MAI-Transcribe-2 | Yes | Yes | none, segment, word | clean, verbatim | Yes |
| ElevenLabs Scribe v2 | Yes | Yes | none, segment, word | clean, verbatim | Yes |
| Groq Whisper large v3 | Yes | No | none, segment, word | verbatim | prompt |
| Groq Whisper large v3 turbo | Yes | No | none, segment, word | verbatim | prompt |
Microsoft requires a Speech resource endpoint and key. OpenTranscribe calls the synchronous Fast Transcription REST endpoint with enhanced mode. ElevenLabs calls /v1/speech-to-text with zero retention enabled by default. That provider option requires an eligible account; operators who deliberately accept provider-side logging may set OT_ELEVENLABS__ZERO_RETENTION=false. Groq uses the OpenAI-compatible audio transcription endpoint and supports either URL passthrough or an explicitly requested proxy upload.
Groq word timestamps are requested together with segment timestamps so word data remains attached to canonical segments instead of producing a structurally incomplete response.
diarization, timestamps, and transcript_style are unset by default, which asks for no particular capability: a request naming only its audio source routes to Groq as readily as to Microsoft, and the response metadata reports what the selected model actually applied. State a capability to require it, and no model lacking it is selected: diarization=true excludes both Groq models, as does transcript_style=clean. Adding strict_capabilities=false downgrades instead of failing, with a requested_capability_not_supported:* warning naming what was dropped.
Pricing metadata lives in config/pricing.yaml. It is informational and deliberately separate from adapter code. Expired or unavailable metadata produces warnings.
French conversations with speaker turns¶
ElevenLabs Scribe v2 is an available option for French audio with speaker diarization. Its transcription documentation lists French and speaker diarization. Availability does not establish an accuracy advantage over Whisper on your recordings: compare the same authorized audio, transcription errors, speaker assignments, latency, and cost before making that claim.
Enable the adapter by supplying OT_ELEVENLABS__API_KEY through your deployment's secret configuration. Keep OT_ELEVENLABS__ZERO_RETENTION=true; the account must be eligible for this option. For the desktop assistant, enable ElevenLabs and store the key through Setup. Use list_transcription_models(configured_only=true) to verify the model is configured. This is not a live transcription or entitlement check.
Request the language and speaker separation explicitly:
{
"source": {"type": "url", "url": "https://example.com/authorized-recording.mp3"},
"provider": "elevenlabs",
"model": "scribe-v2",
"routing_policy": "fixed",
"language": "fr",
"diarization": true,
"timestamps": "segment",
"transcript_style": "verbatim",
"strict_capabilities": true,
"allow_fallback": false,
"result_mode": "auto"
}
The example URL is a placeholder. Supply your authorized HTTPS audio URL only to the tool, keeping signed URLs out of files and logs. If the number of speakers is known, optionally supply speaker_count_hint; otherwise let the model estimate it.
Render the canonical segments as timestamped speaker turns. Their speaker labels remain stable within a transcript, including when a speaker returns: SPEAKER_01, SPEAKER_02, then SPEAKER_01 again. The flat text field does not preserve speaker turns. These labels distinguish voices; they do not identify people. Report missing speaker labels instead of inventing them. Transcription and diarization happen together from the audio; a flat transcript alone cannot recover acoustic speaker boundaries.
Both Groq Whisper models are incompatible with diarization=true. Strict routing rejects them before a provider call. Disabling strict capability checking would drop that requirement with a warning, so keep it enabled when speaker turns matter.