Open-Source Voice AI That Plugs Into Your Asterisk Switch

What it does AVA is a Dockerized voice agent that bridges Asterisk/FreePBX to cloud AI services over Audiosocket/RTP. It turns standard phone calls into streaming conversations with LLMs, managing speech-to-text, reasoning, and text-to-speech in real time. An Admin UI and setup wizard generate dialplan snippets so you can route calls through AI contexts without writing custom AGI scripts.

The interesting bit The project treats telephony as a transport layer rather than an afterthought: μ-law audio at 8 kHz passes straight to compatible providers like xAI Grok with no resampling, and the modular pipeline lets you swap STT, LLM, and TTS components independently instead of buying an all-in-one black box.

Key highlights

• Supports five “full-agent” realtime providers (OpenAI Realtime, Google Live, xAI Grok, Deepgram, ElevenLabs Agent) plus modular STT/LLM/TTS pipelines
• Multi-instance provider support isolates credentials for multi-tenant setups
• Local inference stack for offline operation: Faster-Whisper, Piper TTS, and Qwen 0.5B
• Browser playback of compact μ-law call recordings without transcoding
• Admin UI with per-component health checks and ~260 inline help tooltips

Caveats

• The default Admin UI password is admin/admin, and the README explicitly warns you to change it immediately before exposing port 3003
• Multi-tenant setups using short aliases like AI_PROVIDER=openai broke in v6.5.2 and now require exact provider instance keys
• xAI Grok sessions carry a hard 30-minute cap from the provider side

Verdict Worth a look if you run Asterisk or FreePBX and need to add AI voice capabilities without replacing your dialplan. Skip it if you’re already committed to a cloud contact-center platform that owns the full stack.