Ava – AI Voice Agent for Traditional Phone Systems(Python+Asterisk/ARI)
Hi HN, I'm the creator of AVA - AI Voice Agent for Asterisk My repo was shared here once before by someone else so I wanted to follow up with the progress since then. https://news.ycombinator.com/item?id=46380399 I've been working with Asterisk/FreePBX systems for years. I wanted to add AI voice capabilities to legacy phone systems without paying per-minute SaaS fees or ripping out the entire telephony stack. So I built AVA, a self-hosted AI voice agent that can integrate into any traditional phone system. While most solutions demand expensive migrations to cloud-only…
What it does
In the maker’s words, at launch
Hi HN, I'm the creator of AVA - AI Voice Agent for Asterisk My repo was shared here once before by someone else so I wanted to follow up with the progress since then. https://news.ycombinator.com/item?id=46380399 I've been working with Asterisk/FreePBX systems for years. I wanted to add AI voice capabilities to legacy phone systems without paying per-minute SaaS fees or ripping out the entire telephony stack. So I built AVA, a self-hosted AI voice agent that can integrate into any traditional phone system. While most solutions demand expensive migrations to cloud-only providers, AVA provides a self-hosted path to connect AI agents to existing phone systems while ensuring data privacy and lowering operational costs AVA is a Dockerized Python app that sits alongside your Asterisk server. It connects via ARI (Asterisk REST Interface) and routes call audio to AI providers — OpenAI Realtime, Deepgram, Google Live API, ElevenLabs, Telnyx, or fully local models (Vosk + llama.cpp + Piper). You can mix and match STT/LLM/TTS in a modular pipeline, or use a single provider end-to-end. Two audio transport paths: We support both AudioSocket (low-latency TCP with TLV framing) and ExternalMedia RTP (UDP, better for NAT). A transport orchestrator auto-negotiates sample rates and codecs between what Asterisk sends on the wire and what each AI provider expects — so you can run 8kHz ulaw from Asterisk into a provider that wants 24kHz linear16 without manual config. Session lifecycle: A typed session store tracks every call from StasisStart through hangup — audio diagnostics, barge-in counts, provider state, conversation turns. Every call is fully observable and debuggable after the fact. Barge-in and VAD were the hardest problems. We use a dual-mode VAD — WebRTC VAD combined with energy-based RMS detection, scored into a single confidence value (40% WebRTC weight, 40% energy ratio, 20% agreement bonus). Frame smoothing prevents single-frame glitches from triggering false interrupts. When barge-in fires, we kill active playback (both streaming and file-based) via ARI, flush provider audio buffers, release conversation gating tokens, and optionally suppress provider output for a configurable window to prevent pre-barge audio from re-queuing. The system supports three interrupt sources: local VAD, Asterisk's native talk detection events, and provider-side interruption signals. The hardest latency challenge was bridging legacy SIP/RTP with modern WebSocket streams. We use a two-container architecture: a lightweight orchestrator for ARI state management and an optional heavier container for local model inference. There are 6 pre-validated golden baseline configs if you just want something working out of the box, plus an Admin UI for visual setup. Try the live demo: (925)-736-6718 Option 5 for Google, 6 for Deepgram, 7 for Openai realtime, 8 for Local hybrid and 9 for Elevenlabs Code is MIT. I'd love feedback on the transport layer (src/core/transport_orchestrator.py) and the VAD tuning (src/core/vad_manager.py).
Does the same job
all alternatives →
- CCCallhackernews.com – The missing voice interface for Hacker News2024 · callhackernews.com · ▲9
Hey there HN! I believe the future of AI communication will be more voice and less text. Low-latency realistic voice interactions are finally becoming feasible. I've built a few voice-first apps on Retell AI using Elevenlabs voices. This one uses Claude Haiku for responses and Mixtral to switch between posts and comments. The AI knows about the top 30 posts and their comments on Hacker News right now. After a Google sign-in you can try it free for 10 minutes. I'd love to hear your thoughts!
- AEAudentic – Enable Voice AI for your website2025 · audentic.io · ▲8
Hey HN! Have been working on Audentic, a platform that simplifies adding voice AI to your website. Think of it as a copy-paste voice assistant that you can embed directly into your site. Setting up voice AI has traditionally been a bit of a hassle, often requiring juggling multiple components like speech recognition, text processing, and text-to-speech systems. With Audentic, we've streamlined this into a more straightforward, end-to-end voice model. While OpenAI's Realtime API has made real-time, multimodal AI interactions more feasible, integrating these capabilities into a website can…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
Launched alongside, March 2026
the whole month →
Switch from ChatGPT to Claude with import memory feature
AI · Mar 2026 · claude.com
