voiceblender

VoiceBlender

A Go voice engine that connects SIP, WebRTC, WhatsApp, WebSocket and other call legs into mixable rooms, with recording, speech AI integrations, a REST API and real-time events.

Features

Legs (call endpoints)

• SIP -- inbound and outbound calls over UDP, TCP and TLS, with digest auth, hold, REFER transfer, early media and session timers
• SIP trunks & registrations -- register to upstream PBXs/carriers or peer with them by static IP, and act as a registrar for SIP clients
• Codecs -- PCMU, PCMA, G.722, Opus, AMR-WB, AMR-NB, with mid-call codec renegotiation
• WebRTC -- browser voice via SDP offer/answer with trickle ICE and ICE restarts
• WhatsApp Business Calling -- inbound and outbound WhatsApp voice calls
• WebSocket legs -- raw PCM over WebSocket for bots and custom clients
• LiveKit -- bridge a LiveKit room, one leg per remote participant
• MoQ (experimental) -- Media-over-QUIC legs over WebTransport
• SIPREC -- session recording server and client (RFC 7865 / RFC 7866)
• Multi-stream SIP calls -- several audio streams in one call, e.g. original and translated audio

Media

• Rooms -- multi-party mixing (mixed-minus-self) at 8, 16 or 48 kHz
• Room bridging -- connect two rooms with configurable direction
• Audio routing matrix -- role-based who-hears-whom for whisper, barge-in and supervisor monitoring
• Audio filters -- per-leg denoise, band-pass, gain and voice effects (robotic, vocoder, pitch)
• Comfort noise -- optional per-room background noise
• DTMF & Real-Time Text -- RFC 4733 telephone-events and T.140 text
• Recording -- per-leg, per-room or multi-track WAV with pause/resume and S3/GCS upload
• Playback -- WAV/MP3 files and telephone tones

AI

• TTS -- ElevenLabs, Google Cloud, AWS Polly, Deepgram, Azure, with speculative preflight synthesis
• STT -- ElevenLabs, Deepgram, Deepgram Flux, Azure, Speechmatics, with turn detection
• AI agents -- ElevenLabs, VAPI, Pipecat, Deepgram, attachable to a leg or a whole room
• Answering machine detection -- human / machine classification with beep detection

Control & observability

• REST API -- full call and room control (OpenAPI spec)
• Webhooks -- signed, retried, deduplicated event delivery
• VSI -- a single WebSocket for events and commands (AsyncAPI spec)
• Prometheus metrics -- at GET /metrics, plus optional pprof

Quick Start

# Build and run
go build -o voiceblender ./cmd/voiceblender
./voiceblender

# Or run directly
go run ./cmd/voiceblender

The REST API listens on :8080 and SIP on 127.0.0.1:5060 by default. All configuration is via environment variables -- see CONFIGURATION.md and the ready-to-edit voiceblender.env.example.

To try calls between two instances, bring up the two-node cluster:

docker compose -f docker/docker-compose.cluster.yml up --build

API Overview

All endpoints live under /v1. The table is a map only -- see API.md for requests, responses and events.

Security: the HTTP API has no built-in authentication. Access is gated only by the ALLOWED_IPS allowlist and network placement -- put an authenticating reverse proxy in front of it before exposing it to untrusted networks.

Use Cases

• Voice AI agents -- put an AI agent on inbound or outbound phone, WhatsApp or browser calls
• Contact center -- conferencing, transfers, supervisor whisper / barge-in, recording with PCI pause
• Outbound dialing -- answering machine detection, TTS announcements, DTMF menus
• Call recording & analytics -- SIPREC recording server with live transcription
• Live translation -- separate rooms for the original and translated audio of one call
• Protocol gateway -- bridge SIP, WebRTC, WhatsApp, LiveKit and WebSocket participants in one room

Your First Call

With VoiceBlender running and a SIP endpoint to call, dial it into a room, speak to it, record it and hang up:

API=http://localhost:8080/v1
JSON='Content-Type: application/json'

# Create a room and start recording it
curl -X POST $API/rooms -H "$JSON" -d '{"id":"demo"}'
curl -X POST $API/rooms/demo/record -H "$JSON" -d '{}'

# Dial a SIP phone; it joins the room once answered
LEG=$(curl -s -X POST $API/legs -H "$JSON" \
  -d '{"type":"sip","to":"sip:alice@192.168.1.100:5060","room_id":"demo"}' | jq -r .id)

# After it answers, say something (needs ELEVENLABS_API_KEY), then hang up
curl -X POST $API/legs/$LEG/tts -H "$JSON" \
  -d '{"text":"Hello from VoiceBlender","voice":"Rachel","provider":"elevenlabs"}'
curl -X DELETE $API/legs/$LEG

Watch the call progress (leg.ringing, leg.connected, tts.finished, leg.disconnected, ...) by setting WEBHOOK_URL or connecting a WebSocket client to ws://localhost:8080/v1/vsi.

Noise Suppression

Every leg can run built-in noise suppression (denoise), so a caller in a car, a café or on a noisy trunk reaches everyone else, the recording, STT and the AI agent already cleaned up.

• Two models, pure Go -- RNNoise via rnnoise-go, or GTCRN (denoise_gtcrn, lower CPU) via gtcrn-go; no cgo, no external service, no per-minute cost
• Native rate -- denoise runs at the call's own 8, 16 or 48 kHz, with no extra resampling
• Per leg or server-wide -- set filters on a leg, or AUDIO_FILTERS=denoise as the default
• Live toggle -- switch it on or off mid-call; suppression ramps in over about a second

Denoise can be chained with bandpass, gain and voice effects -- see Audio filters.

Typical Workflow

1. Configure a webhook        WEBHOOK_URL env var, or per leg / room
2. Receive inbound call       --> webhook: leg.ringing {leg_id, from, to}
3. Answer                     POST /v1/legs/{id}/answer
4. Create a room              POST /v1/rooms
5. Add legs to room           POST /v1/rooms/{id}/legs
6. Attach AI agent            POST /v1/legs/{id}/agent/{provider}
7. Start recording            POST /v1/rooms/{id}/record
8. Hang up                    DELETE /v1/legs/{id}

Documentation

• API.md -- REST API, webhooks and event reference
• CONFIGURATION.md -- every environment variable
• TESTING.md -- unit, integration and benchmark tests
• openapi.yaml / asyncapi.yaml -- machine-readable specs
• voiceblender.org/docs -- online documentation

Examples

Complete reference applications built on VoiceBlender live in VoiceBlender/example-applications:

Testing

go test ./internal/...                                              # unit tests
go test -tags integration -v -timeout 60s ./tests/integration/      # integration tests

See TESTING.md for setup and benchmarks.

Contributing

CLAUDE.md is the checklist for every change -- formatting, spec regeneration, docs and tests. It applies to human contributors and any coding agent alike, not just Claude. CI runs:

gofmt -l .
go vet ./...
go test ./internal/... -count=1 -timeout=60s

AI-Assisted Development

This project was developed with the assistance of Claude Code, Anthropic's AI coding assistant.

License

See LICENSE file.