Speech Transcribe

Use the native audio_transcribe tool when the user asks to transcribe, caption, diarize, timestamp, or identify speakers in an audio or video clip.

Workflow

1. Call audio_transcribe with action: "list" if you need provider readiness or the user asks what is configured.
2. Pass audio as a current attachment filename/ref, /workspace path, /discord-media-cache path, /uploaded-media-cache path, or HTTPS media URL.
3. Use provider: "auto" unless the user asks for openai, deepgram, or assemblyai, or unless diarization is required. Auto mode honors the tenant skills.speechToText.defaultProvider config when it is set. Prefer Deepgram or AssemblyAI for speaker labels.
4. Pass language only when the user gives a known language. Omit it for provider language detection.
5. Set diarization: true when the user asks for speaker labels.
6. Set timestamps to word, segment, or none based on the request.
7. Return the structured result fields that matter: transcript text, provider, detected language, duration, cost, warnings, and artifact paths.

The native tool owns provider credentials, provider fallback, output schema, long-audio chunking for local and remote OpenAI uploads when ffmpeg/ffprobe are available, transcript artifact persistence, and usage-cost accounting.

Output Contract

The tool returns JSON shaped like:

{
  "text": "Transcript text",
  "segments": [{ "start": 0, "end": 1.2, "speaker": "speaker_0", "text": "..." }],
  "language": "en",
  "provider": "deepgram",
  "duration_sec": 12.3,
  "cost_usd": 0.001
}

Transcript text and segment JSON are also persisted as private workspace artifacts. Treat transcripts as operator-private until the user explicitly asks to share, post, email, or publish them.