Transcription Built Into the Call Path: Plans, Live Streams and Wrap-Up Notes

Transcription is usually bolted on after the fact: record the call, upload it somewhere, run ASR, reconcile the text back to the call id by filename. It works until someone asks “what did the customer say during the call?”

RustPBX puts transcription in the call path, per call, with the policy you specify.

Per-Call Plans, Not a Global Switch

Transcription is configured as a plan: provider, language, and whether to auto-start — decided per call, not globally:

Auto-start for support queues; off for internal extensions
Language per IVR selection (the customer picked Spanish at the menu)
Provider per plan (cheap batch ASR for volume, premium streaming for VIP)

This runs off the platform’s ASR integration, so switching providers is configuration, not a migration.

Live Streams for Agent Assist

During the call, live transcription can be consumed over an SSE endpoint — the stream an agent-assist UI subscribes to for real-time hints, or that a supervisor watches for escalation keywords. Because the stream is tied to the call, no audio leaves the PBX to make it work.

The CC desk uses this for the “live coach” surface; the same stream feeds external AI tooling.

Wrap-Up and CDR

When the call ends, results land on the call record:

Call record detail — the same record carries recording, trace and media quality

has_transcript / transcript_status / transcript_language on the CDR
The transcript itself attached as a call artifact, retrievable via the session-artifacts API
Searchable alongside the recording segments (per-segment recording ids and metadata already correlate text to audio)

Quality reviewers get call + audio + text in one place, keyed the same way, without filename archaeology.

How It Fits the Bot Story

Transcription is the assist layer under the AI-voice story:

Step IVR / realtime bot handles the conversation (see AI voice bots with handoff)
Transcription records it and streams it to anyone watching
Post-call summaries and QA searches run on the transcript attached to the record

One call, one session id, one CDR — bot phase and human phase included.

Configuration Notes

Enable the transcript addon for the offline/batch pipeline; per-call plans drive live behavior
Recording and transcription are independent — you can transcribe without storing audio, subject to policy
Language detection: set it per plan; don’t rely on a single global language for mixed-audience queues

Guides: Diagnostics → transcription fields and the contact-center reports that consume it.