Transcription

transcription: sends every finished recording to a speech-to-text server and keeps the text with the call. It is the transcripts column the call-playback ecosystem (Scanner Radio’s live transcriptions, trunk-recorder’s transcription plugins) has grown, with a keyword hook into alerts.

GopherTrunk runs no model itself. It speaks the OpenAI audio-transcriptions request shape — multipart file + model + language (+ optional prompt), JSON {"text": …} back — which is what these all expose:

Server	url
OpenAI	https://api.openai.com/v1/audio/transcriptions + api_key
whisper.cpp server	http://host:8080/inference
faster-whisper-server / Speaches	http://host:8000/v1/audio/transcriptions
LocalAI	http://host:8080/v1/audio/transcriptions

transcription:
  enabled: true
  url: "http://127.0.0.1:8080/inference"
  language: "en"
  prompt: "Engine 1, Medic 7, Main Street"
  systems: ["Metro P25"]
  min_duration_ms: 1000

What happens to the text

• Stored on the call (call_log.transcript): the History panel shows a Transcript column and the full text in the call detail; the q search matches transcript text; GET /api/v1/calls/history?format=csv exports it. A multi-over call’s segments append in order.
• Published as the call.transcript event (system, group_id, source_id, call_started_at, segment, audio_path, text).
• Alertable: an alerts: rule with on: [call.transcript] and keywords: ["shots fired", "structure fire"] fires only when the text contains one of the words — with attach_audio: true the recording rides along to Discord / Telegram.

Audio sent

By default the recording (WAV or FLAC) is converted to 16 kHz mono 16-bit WAV (upload_format: wav16k), the one input every Whisper server accepts — whisper.cpp refuses anything else. original uploads the file as recorded. Encrypted calls are skipped (skip_encrypted, default on) and so are recordings shorter than min_duration_ms (default 1 s).

Operating it

GET /api/v1/transcription reports sent / failed / skipped / dropped counts, the mean request latency, the audio seconds transcribed and the last transcript. Requests run on workers goroutines (default 1 — a local server serialises anyway and a hosted one rate-limits) behind a bounded queue; a dead server never stalls the recorder. timeout (default 60 s) bounds one request.

A local whisper.cpp server on a small box keeps up with a busy system with the base.en or small.en model; prompt with your unit names and street names measurably improves proper nouns.