active-call is a standalone Rust crate for building AI Voice Agents. It provides high-performance infrastructure bridging AI models with real-world telephony and web communications.
📖 Documentation → English | 中文 | API Reference
- SIP (Telephony): UDP, TCP, TLS (SIPS), WebSocket. Register as extension to FreeSWITCH / Asterisk / RustPBX, or handle direct SIP calls. PSTN via Twilio and Telnyx.
- WebRTC: Browser-to-agent SRTP. (Requires HTTPS or 127.0.0.1)
- Voice over WebSocket: Push raw PCM/encoded audio, receive real-time events.
- Traditional Pipeline: VAD → ASR → LLM → TTS. Supports OpenAI, Aliyun, Azure, Tencent, Deepgram and more.
- Realtime Streaming: Native OpenAI/Azure Realtime API — full-duplex, ultra-low latency.
- Audio Processing: Built-in noise reduction (nnnoiseless), WebRTC AGC2 automatic gain control, configurable interruption strategy with filler-word filtering.
Define personas, scenes, and flows in Markdown files with rich config:
---
asr:
provider: "sensevoice"
tts:
provider: "supertonic"
speaker: "F1"
llm:
provider: "openai"
model: "${OPENAI_MODEL}"
apiKey: "${OPENAI_API_KEY}"
features: ["intent_clarification", "emotion_resonance"]
dtmf:
"0": { action: "hangup" }
dtmfCollectors:
order_number:
description: "Collect order number"
maxDigits: 6
finishKey: "#"
timeout: 15
interruption:
strategy: "both"
fillerWordFilter: true
agc: {}
denoise: true
ringbackDetection:
enabled: true
sip:
extractHeaders: ["X-Customer-Id"]
posthook:
url: "https://api.example.com/webhook"
summary: "detailed"
---
# Scene: greeting
<dtmf digit="1" action="goto" scene="tech_support" />
<dtmf digit="2" action="transfer" target="sip:billing@domain.com" />
You are a friendly AI for {{ company_name }}. Greet the caller warmly.
To look up info: <http url="https://api.example.com/user/{{ sip["X-Customer-Id"] }}" />
To collect order: <collect type="order_number" var="order_id" prompt="Please enter your 6-digit order number." />
To send metadata: <message body="order={{ order_id }}" />
# Scene: tech_support
How can I help with your system? I can transfer you: <refer to="sip:human@domain.com" />
To search knowledge base: <http url="https://kb.example.com/search" body='{"q":"{{ caller }}"}' method="POST" />${VAR}= environment variables (config-time).{{var}}= runtime variables (per-call).- Built-in vars:
{{ session_id }},{{ call_type }},{{ caller }},{{ callee }},{{ start_time }} - SIP header access:
{{ sip["X-Header-Name"] }} - Scene-level
<collect>triggers DTMF digit collection; results stored as{{ var }}. <set_var key="k" value="v" />and<http url="..." />extend LLM capabilities at runtime.
Run ASR and TTS locally — no cloud API required:
- Offline ASR: SenseVoice — zh, en, ja, ko, yue
- Offline TTS: Supertonic — en, ko, es, pt, fr
# Download models
docker run --rm -v $(pwd)/data/models:/models \
ghcr.io/miuda-ai/active-call:latest \
--download-models all --models-dir /models --exit-after-download
# Run with offline models
docker run -d --net host \
-v $(pwd)/data/models:/app/models \
-v $(pwd)/config:/app/config \
ghcr.io/miuda-ai/active-call:latestMainland China: Add
-e HF_ENDPOINT=https://hf-mirror.comto use the HuggingFace mirror.
| VAD Engine | Time (60s audio) | RTF | Note |
|---|---|---|---|
| TinySilero | ~60 ms | 0.0010 | >2.5× faster ONNX |
| ONNX Silero | ~158 ms | 0.0026 | Standard baseline |
| WebRTC VAD | ~3 ms | 0.00005 | Legacy |
Codec support: PCM16, G.711 (PCMU/PCMA), G.722, Opus.
| Feature | Description |
|---|---|
| Denoise | nnnoiseless-based real-time noise reduction |
| AGC | WebRTC AGC2 adaptive gain control |
| Ambiance | Background audio mixing with ducking |
| Interruption | Configurable VAD/ASR/None strategy + fade-out |
- Bridge/Unbridge: Bidirectional audio patching between two active calls (media only, no SIP signaling).
- Pause/Resume: Pause/resume server-side TTS or file playback.
- Mute/Unmute: Mute/unmute specific audio tracks.
- Graceful Shutdown: Waits for active calls to finish naturally before exit.
# Webhook handler
./active-call --handler https://example.com/webhook
# Playbook handler
./active-call --handler config/playbook/greeting.md
# Outbound SIP call
./active-call --call sip:1001@127.0.0.1:5060 --handler greeting.md
# With external IP and codecs
./active-call --handler default.md --external-ip 1.2.3.4 --codecs pcmu,pcma,opusdocker run -d --net host \
--name active-call \
-v $(pwd)/config.toml:/app/config.toml:ro \
-v $(pwd)/config:/app/config \
ghcr.io/miuda-ai/active-call:latest[handler]
type = "playbook"
default = "greeting.md"
[[handler.rules]]
caller = "^\\+1\\d{10}$"
callee = "^sip:support@.*"
playbook = "support.md"
[[handler.rules]]
caller = "^\\+86\\d+"
playbook = "chinese.md"tls_port = 5061
tls_cert_file = "./certs/cert.pem"
tls_key_file = "./certs/key.pem"
enable_srtp = true| Feature | Cargo flag | Description |
|---|---|---|
| offline (default) | --features offline |
SenseVoice ASR + Supertonic TTS |
| cross | --features cross |
offline + AWS-LC bindgen (cross-compile) |
| ringback-detection | --features ringback-detection |
Ringback tone detection |
| opus | --features opus |
Opus codec support |
| not_vad | --no-default-features |
Exclude VAD (for external VAD) |
cargo build --release
cargo build --release --features ringback-detection,opus# rustc may require extra stack when compiling certain crates
export RUST_MIN_STACK=16777216
cargo build --releaseOr add to your shell profile:
echo 'export RUST_MIN_STACK=16777216' >> ~/.bashrc# OpenAI / Azure
OPENAI_API_KEY=sk-...
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
# Aliyun DashScope
DASHSCOPE_API_KEY=sk-...
# Tencent Cloud
TENCENT_APPID=...
TENCENT_SECRET_ID=...
TENCENT_SECRET_KEY=...
# Deepgram
DEEPGRAM_API_KEY=...
# 60db
SIXTYDB_API_KEY=...
SIXTYDB_VOICE_ID=... # UUID from GET https://api.60db.ai/voices
# Offline models
OFFLINE_MODELS_DIR=/path/to/models60db is a hosted speech API. Select it in a playbook using a voice UUID from
GET https://api.60db.ai/voices:
tts:
provider: "60db"
apiKey: "${SIXTYDB_API_KEY}"
speaker: "${SIXTYDB_VOICE_ID}"
samplerate: 16000
codec: "pcm"
speed: 1.0The provider supports mono PCM16 at 16000, 24000 or 48000 Hz, a speaking rate
from 0.5 to 2.0, and an optional language hint. It sends each text command over
HTTP in order and accepts JSON or NDJSON audio responses. For streaming tracks,
completion is emitted after all accepted commands finish. Dropping the track
cancels its HTTP response. endpoint can override the full synthesis URL.
| Method | Path | Description |
|---|---|---|
| GET | /call |
WebSocket call handler |
| GET | /call/webrtc |
WebRTC call handler |
| GET | /call/sip |
SIP call handler |
| GET | /list |
List all active calls |
| GET | /kill/{id} |
Force-terminate an active call |
| GET | /events/{id} |
SSE event stream for a call |
| POST | /command/{id} |
Send command to an active call |
| POST | /precache |
Pre-generate TTS audio to cache |
| GET | /iceservers |
ICE server configuration for WebRTC |
| GET | /api/playbooks |
List available playbooks |
| GET | /api/playbooks/{name} |
Get playbook content |
| POST | /api/playbooks/{name} |
Save/update a playbook |
| POST | /api/playbook/run |
Associate a playbook with a new session |
| GET | /api/records |
List call event records |
| GET | / |
Serve built-in web client |
Full command/event reference → API Documentation
- Go: rustpbxgo — Official Go client
| Language | Links |
|---|---|
| English | Docs Hub · API Reference · Config Guide · Playbook Tutorial · Advanced Features |
| 中文 | 文档中心 · API 文档 · 配置指南 · Playbook 教程 · 高级特性 |
MIT — see LICENSE
