Voice Transcribe

Convert a voice note into text so the normal chat pipeline can handle it. This runs as a pre-processing step on inbound audio, before routing.

Pipeline

1. Receive — the channel adapter saves the audio attachment to a temp path.
2. Transcribe — run speech-to-text (preference order):
   • OpenAI Whisper API if OPENAI_API_KEY is set (fast, no local deps):

     curl -s https://api.openai.com/v1/audio/transcriptions \
       -H "Authorization: Bearer $OPENAI_API_KEY" \
       -F model=whisper-1 -F file=@note.ogg
     

   • Local faster-whisper if no key is configured and CPU/GPU allows.
3. Inject — replace the message body with the transcript, tag it [voice], and hand it to the router as if the user had typed it.
4. Clean up — delete the temp audio file after transcription.

Notes

• WhatsApp/Telegram voice notes are usually .ogg (Opus). Whisper accepts it directly; faster-whisper may need ffmpeg installed.
• Keep the original duration in metadata so the agent can say "got your 0:42 voice note" when useful.
• Transcription is read-only — no capability gate needed. The reply still goes out through the normal gated chat.send path.
• Never persist raw audio beyond the transcription step unless the user opts in.