Active Call

active-call is a standalone Rust crate designed for building AI Voice Agents. It provides a high-performance infrastructure for bridging AI models with real-word telephony and web communications.

Key Capabilities

1. Multi-Protocol Audio Gateway

active-call supports a wide range of communication protocols, ensuring compatibility with both legacy and modern systems:

• SIP (Telephony): Supports standard SIP signaling. Can act as a SIP Client/Extension to register with PBX systems like FreeSWITCH or Asterisk, or handle direct SIP incoming/outgoing calls.
• WebRTC: Direct browser-to-agent communication with low-latency SRTP.
• Voice over WebSocket: A highly flexible API for custom integrations. Push raw PCM/encoded audio over WebSocket and receive real-time events.

2. Dual-Engine Dialogue Support

Choose the pipeline that fits your latency and cost requirements:

• Traditional Serial Pipeline: Integrated VAD → ASR → LLM → TTS flow. Supports various providers (OpenAI, Aliyun, Azure, Tencent) with optimized buffering and endpoints.
• Realtime Streaming Pipeline: Native support for OpenAI/Azure Realtime API. True full-duplex conversational AI with ultra-low latency, server-side VAD, and emotional nuance.

3. Playbook: The Best Practice for Dialogue

The Playbook system is our recommended way to build complex, stateful voice agents:

• Markdown-Driven: Define personas, instructions, and flows in readable Markdown files.
• Stateful Scenes: Manage conversation stages with easy transitions (Scene switching).
• Tool Integration: Built-in support for DTMF, SIP Refer (Transfer), and custom Function Calling.
• Advanced Interaction: Smart interruptions, filler word filtering, background ambiance, and automated post-call summaries via Webhooks.

4. High-Performance Media Core

• Low-Latency VAD: Includes TinySilero (optimized Rust implementation), significantly faster than standard ONNX models.
• Flexible Processing Chain: Easily add noise reduction, echo cancellation, or custom audio processors.
• Codec Support: PCM16, G.711 (PCMU/PCMA), G.722, and Opus.

Protocol Flexibility

Voice over WebSocket

For developers who need full control, active-call provides a raw audio-over-websocket interface. This is ideal for custom web apps or integration with existing AI pipelines where you want to handle the audio stream manually.

SIP PBX Integration

active-call can be integrated into existing corporate telephony:

• As an Extension: Register active-call to your FreeSWITCH or Asterisk PBX like any other VoIP phone. AI agents can then receive calls from internal extensions or external trunks.
• As a Trunk: Handle incoming SIP traffic directly from carriers.

Playbook Demo

VAD Performance

Benchmarked on 60 seconds of 16kHz audio (Release mode):

VAD Engine	Implementation	Time (60s)	RTF (Ratio)	Note
TinySilero	Rust (Optimized)	~60.0 ms	0.0010	>2.5x faster than ONNX
ONNX Silero	ONNX Runtime	~158.3 ms	0.0026	Standard baseline
WebRTC VAD	C/C++ (Bind)	~3.1 ms	0.00005	Legacy, less accurate

API Documentation

For detailed information on REST endpoints and WebSocket protocols, please refer to the API Documentation.

SDKs

• Go SDK: rustpbxgo - Official Go client for building voice applications with active-call.

Dialogue Features

active-call provides a rich set of features for building natural and responsive voice agents:

• Scene Management (Playbook):

  • Markdown-based configuration: Define agent behavior using simple Markdown files.
  • Multiple Scenes: Split long conversations into manageable stages (Scenes) using # Scene: name headers.
  • Dynamic Variables: Supports {{ variable }} syntax using minijinja for personalized prompts.
  • LLM Integration: Streaming responses with configurable prompts and tool-like tags.
  • Background Ambiance: Supports looping background audio (e.g., office noise) with auto-ducking during conversation.
  • DTMF Support: Standard IVR functionality. Define global or scene-specific keypress actions (jump to scene, transfer call, or hang up).
  • Pre-recorded Audio: Play fixed audio files (.wav/.pcm) by adding <play file="path/to/audio.wav" /> at the beginning of a scene.
  • Post-hook & Summary: Automatically generate conversation summaries and send them to a Webhook URL after the call ends. Supports multiple summary templates:
    • short: One or two sentence summary.
    • detailed: Comprehensive summary with key points and decisions.
    • intent: Extracted user intent.
    • json: Structured JSON summary.
    • custom: Fully custom summary prompt.

• Advanced Voice Interaction:

  • Smart Interruption: Multiple strategies (vad, asr, or both), filler word filtering, and protection periods.
  • Graceful Interruption: Supports audio fade-out instead of hard cuts when the user starts speaking.
  • EOU (End of Utterance) Optimization: Starts LLM inference as soon as silence is detected, before ASR final results are ready.

Playbook Configuration

Playbooks are defined in Markdown files with YAML frontmatter. The frontmatter configures the voice capabilities, while the body defines the agent's persona and instructions.

Example

---
asr:
  provider: "openai"
tts:
  provider: "openai"
llm:
  provider: "openai"
  model: "gpt-4-turbo"
dtmf:
  "0": { action: "hangup" }
posthook:
  url: "https://api.example.com/webhook"
  summary: "detailed"
  includeHistory: true
---

# Scene: greeting
<dtmf digit="1" action="goto" scene="tech_support" />

You are a friendly AI for {{ company_name }}. 
Greet the caller and ask if they need technical support (Press 1) or billing help.

# Scene: tech_support
You are now in tech support. How can I help with your system?
To speak to a human, I can transfer you: <refer to="sip:[email protected]" />

Configuration Reference

Section	Field	Description
asr	provider	Provider name (e.g., aliyun, openai, tencent).
tts	provider	Provider name.
llm	provider	LLM Provider.
	model	Model name (e.g., gpt-4, qwen-plus).
dtmf	digit	Mapping of keys (0-9, *, #) to actions.
	action	goto (scene), transfer (SIP), or hangup.
interruption	strategy	both, vad, asr, or none.
	fillerWordFilter	Enable filler word filtering (true/false).
vad	provider	VAD provider (e.g., silero).
realtime	provider	openai or azure. Enable low-latency streaming pipeline.
	model	Specific realtime model (e.g., gpt-4o-realtime).
ambiance	path	Path to background audio file.
	duckLevel	Volume level when agent is speaking (0.0-1.0).
recorder	recorderFile	Path template for recording (e.g., call_{id}.wav).
denoise	-	Enable/Disable noise suppression (true/false).
greeting	-	Initial greeting message.

Docker Usage

Pull the Image

docker pull ghcr.io/restsend/active-call:latest

Configuration

Copy the example config and customize it:

cp active-call.example.toml config.toml

Run with Docker

docker run -d \
  --name active-call \
  -p 8080:8080 \
  -p 13050:13050/udp \
  -v $(pwd)/config.toml:/app/config.toml:ro \
  -v $(pwd)/config:/app/config \
  ghcr.io/restsend/active-call:latest

Environment Variables

If you have API keys, save them in an .env file:

TENCENT_APPID=your_app_id
TENCENT_SECRET_ID=your_secret_id
TENCENT_SECRET_KEY=your_secret_key
DASHSCOPE_API_KEY=your_dashscope_api_key

And mount it in the container like :

  -v $(pwd)/.env:/app/.env \

Port Range

Use small range 20000-20100 for local development, bigger range like 20000-30000, or host network for production.

License

This project is licensed under the MIT License.