Agent Voice Response (commonly abbreviated AVR) is a conversational AI platform that integrates with the open-source PBX Asterisk to replace traditional IVR (Interactive Voice Response) systems using AI-driven voice interactions. It acts as a voicebot, transcribing speech, generating responses through large language models (LLMs), and synthesizing speech back to callers.
AVR offers a Docker-based deployment enabling organizations to integrate automatic speech recognition (ASR), large language models (LLM), text-to-speech (TTS), and direct speech-to-speech (STS) in voice calling systems. It supports a flexible, provider-agnostic architecture, allowing integration with both cloud services and local/open-source providers.
Examples of supported providers include:
- Cloud services: Google, Deepgram, OpenAI, Anthropic
- Local / open-source options: Vosk ASR, Kokoro TTS, CoquiTTS, Ollama LLM
A key part of AVR’s design is its intelligent audio handling:
- Voice Activity Detection (VAD): Ensures responsive and natural turn-taking during calls, allowing users to interrupt the agent smoothly
- Noise Handling & Ambient Sounds: Enables background noise management for realistic or adaptive environments
- Speech-to-Speech (STS) Mode: Delivers real-time voice-to-voice conversations without intermediate text processing
- Modular AI Components: Swappable ASR, LLM, TTS, and STS blocks via HTTP APIs, including support for local providers like Vosk, Kokoro, CoquiTTS, and Ollama
- Voice Activity Detection (VAD): Enables natural interruption and fast turn-taking
- Noise & Ambient Sound Control: Background audio simulation and noise reduction
- Dockerized Microservices: Each core service runs isolated in containers for scalability
- Support for Function-Calling: Integration with OpenAI and Anthropic assistive features
- Speech-to-Speech Mode (STS): Direct voice transformation using providers like OpenAI Realtime, Ultravox, or Deepgram
- Multi-language & Custom Voices: Configurable voice personalities and multilingual support
- Core Service: Orchestrates voice input/output via Asterisk and AudioSocket
- ASR Service: Transcribes caller audio into text (supports cloud and local providers like Vosk)
- LLM Service: Generates conversational responses (supports OpenAI, Anthropic, Ollama)
- TTS Service: Converts responses back to speech (supports cloud and local engines like Kokoro, CoquiTTS)
- STS Service: (Optional) Provides direct speech-to-speech mode for ultra-low latency
- VAD & Noise Engine: Handles voice activity detection and background noise simulation
- Web GUI: (Under development) for deployment management
- Asterisk Integration: Communication via SIP and AudioSocket
The avr-infra GitHub repository provides Docker Compose templates integrating AVR Core, Asterisk, and configurable ASR/LLM/TTS/STS stacks.
Supported combinations include:
- Cloud providers: Deepgram + OpenAI, Google + OpenRouter, ElevenLabs + Anthropic
- Local / open-source providers: Vosk + Ollama + CoquiTTS
- Full voice pipelines: OpenAI Realtime, Ultravox
¶ Use and Community
- Installation: AVR can be deployed on any system running Docker and Asterisk (v18+ with AudioSocket) or on Asterisk-based platforms such as FreePBX, VitalPBX or Vicidial. Users register a SIP client and call a defined extension to interact with the AI agent.
- Licensing: AVR is free for personal and commercial use. The core AVR source remains proprietary, while the infrastructure and integration repositories are fully open-source.
- Contributions: The community is active on Discord, and the project accepts donations via Ko-fi.
For a complete list of AVR repositories, please refer to the GitHub page: https://github.com/orgs/agentvoiceresponse/repositories.
You can search the repositories by keywords such as ASR, LLM, TTS, STS, etc., to find the specific modules and integrations you need.