AVA-AI-Voice-Agent-for-Asterisk

Voice now belongs to agents. Configure one provider, create multiple agents that share it — each with its own voice.

✨ Highlights

🎙️ Per-agent voice selection

• Provider-aware voice picker in the Agent form: dropdowns of validated catalogs for OpenAI Realtime (10 GA voices), Google Live (30 prebuilt voices), and Deepgram Voice Agent (Aura models) — plus suggestions with custom cloned-voice IDs for xAI Grok. ElevenLabs Agent voices remain managed on the ElevenLabs platform; pipeline TTS voices are unchanged (per-agent pipeline voice is planned).
• Safe by default. The provider-level voice becomes the Default Voice — every agent without its own voice behaves exactly as before this release. Unrecognized values (including stale text from the old informational field) log a warning and fall back to the provider default — a bad voice value never fails a call.
• Observable end-to-end. Every call logs the resolved voice and its source (override / agent / provider-default), and Call History shows it per call — e.g. "Voice: marin (from agent)" — so you can always verify what a call actually used.
• Instant. Agent voice changes apply on the next call — no engine restart.

🧰 Under the hood

• Single-source voice catalog (src/utils/voice_catalog.py) shared by the engine's validation, the new GET /api/config/providers/meta endpoint, and the Agent form — verified complete against official OpenAI / xAI / Google / Deepgram documentation.
• Call History gains voice / voice_source (additive, nullable — no migration steps).
• Optional voice: key on YAML contexts for headless installs, with identical semantics.

⬆️ Upgrading

Standard upgrade — no migration steps, no config changes required. Existing agents have no voice set and keep using their provider's voice.

git fetch --tags && git checkout v7.3.0
docker compose build admin_ui && docker compose up -d

Or use the Admin UI's Updates page.

Docs: Voice Selection guide · validation matrix (7 live PSTN calls, including the per-agent override validated by ear)

Thanks to @foytech, whose PR #497 seeded this feature and whose production testing shaped it — co-authored on the core commits.