Anonymous Outbound Voice Agent

Place real phone calls with an AI voice agent without revealing your real name.

The agent dials real numbers through Vobiz (telephony / SIP carrier) connected to LiveKit Cloud, speaks and listens via Sarvam AI (Telugu by default), and reasons with Grok through OpenRouter. You get a live terminal dashboard, optional browser listen-in, and transcripts — while the bot only uses a public alias.

---

Features

• Anonymous identity — agent uses a configurable alias (e.g. “Alex”), never your real name, phone, or email
• Outbound phone calls — Vobiz SIP trunk + LiveKit Cloud realtime
• Multilingual voice — Sarvam STT/TTS with Telugu (te-IN) default; English and other Indian languages supported
• Low-latency LLM — Grok via OpenRouter with light reasoning tuned for natural phone speech
• Live call monitor — dialing / ringing / answered / ended, timers, and live transcript
• Listen-only mode — hear the call on speakers or in the browser without joining the conversation
• Noise-resistant barge-in — short coughs and noise are less likely to cut the agent off mid-sentence
• Auto hang-up — silence / max duration / goodbye detection

---

Stack

Layer	Provider
Telephony (SIP)	Vobiz — outbound SIP trunk for real phone calls
Realtime media	LiveKit Cloud (rooms, agents, SIP participant bridge)
STT	Sarvam AI (saaras:v3, auto-detect incl. Telugu)
LLM	OpenRouter → Grok (x-ai/grok-4.3 by default)
TTS	Sarvam AI (bulbul:v3, default Telugu te-IN)
VAD	Silero (via LiveKit agents)

Telephony partner: Vobiz

This project places real PSTN/mobile calls using a Vobiz outbound SIP trunk registered in LiveKit:

1. Create a trunk in the Vobiz dashboard (or equivalent Vobiz SIP credentials).
2. Add that trunk as an outbound SIP trunk in LiveKit Cloud (SIP → Trunks).
3. Copy the LiveKit trunk id (ST_…) into .env as SIP_OUTBOUND_TRUNK_ID.

When you run make_call.py, LiveKit creates a SIP participant and Vobiz dials the destination number. Call states (dialing, ringing, active, hangup) come from LiveKit SIP attributes over that trunk.

---

Prerequisites

• Python 3.11+ recommended
• Accounts / keys:
  • LiveKit Cloud project (LIVEKIT_URL, API key + secret)
  • Vobiz telephony account + outbound SIP trunk linked in LiveKit → SIP_OUTBOUND_TRUNK_ID
  • Sarvam AI API key
  • OpenRouter API key
• Windows, macOS, or Linux (Windows users should run the included local-inference patch)

---

Quick start

1. Clone and install

git clone https://github.com/Iamnanisyam/voice-agent.git
cd voice-agent

python -m venv venv

# Windows
.\venv\Scripts\Activate.ps1

# macOS / Linux
# source venv/bin/activate

pip install -r requirements.txt

# Windows only — avoid livekit-local-inference ACCESS_VIOLATION crash
python patch_local_inference.py

# Download model assets (Silero VAD, etc.)
python agent.py download-files

2. Configure secrets

copy .env.example .env

Edit .env and set at least:

Variable	Description
LIVEKIT_URL	e.g. wss://your-project.livekit.cloud
LIVEKIT_API_KEY	LiveKit API key
LIVEKIT_API_SECRET	LiveKit API secret
SARVAM_API_KEY	Sarvam STT + TTS
OPENROUTER_API_KEY	OpenRouter key for Grok
OPENROUTER_MODEL	Default: x-ai/grok-4.3
SIP_OUTBOUND_TRUNK_ID	LiveKit id of your Vobiz outbound SIP trunk (ST_…)

Optional identity (privacy):

AGENT_PUBLIC_NAME=Alex
AGENT_REPRESENTS_AS=a private client
AGENT_LANGUAGE=te-IN
AGENT_SPEAKER=shubh

Never commit .env. It is listed in .gitignore.

3. Start the agent worker

Keep this terminal running:

.\venv\Scripts\Activate.ps1
python agent.py dev

4. Place a call (second terminal)

.\venv\Scripts\Activate.ps1
python make_call.py `
  --to +919876543210 `
  --purpose "Confirm whether the meeting tomorrow at 3pm still works" `
  --callee "Ravi" `
  --context "Meeting is optional; if busy, suggest Friday afternoon"

By default the script watches the call live in that terminal.

---

How to use

Call status dashboard

While a call is active, make_call.py and the worker show:

Status	Meaning
INITIATING	Dispatch accepted, agent joining
DIALING…	SIP INVITE via Vobiz trunk
RINGING…	Remote phone ringing
LIFTED (active)	Callee answered
BUSY / REJECTED / NO ANSWER	Failed outcomes
ENDED	Hangup

Also shown: elapsed time since dial, talk time since answer, and a LIVE TRANSCRIPT (THEM = callee, AGENT = bot).

Dispatch only (no watch loop):

python make_call.py --to +919876543210 --purpose "…" --no-watch

Useful flags

Flag	Meaning
--to	E.164 number (required), e.g. +919876543210
--purpose	Why the agent is calling (required for a useful call)
--callee	Name of the person being called
--context	Extra background for the agent
--public-name	Alias used on the call (not your real name)
--represents-as	Who the agent says it represents
--language	TTS language (te-IN, en-IN, hi-IN, …)
--speaker	Sarvam voice (shubh, priya, …)
--no-watch	Exit after dispatch instead of live monitoring

Live listen (optional third terminal / browser)

While a call is active:

# Speakers only (subscribe-only, no mic)
python listen.py
# or pin a room from make_call output:
python listen.py --room call-91xxxxxxxxxx-abc

# Browser player + optional recording
python listen.py --browser

You join as a listen-only participant and hear agent + callee audio.

Console test (no phone)

Verify STT / LLM / TTS without dialing:

python agent.py console

Validate a Sarvam / ElevenLabs voice id

python validate_voice.py --speaker shubh

Dispatch via LiveKit CLI (alternative)

lk dispatch create \
  --new-room \
  --agent-name anonymous-caller \
  --metadata "{\"phone_number\": \"+919876543210\", \"purpose\": \"Confirm the appointment\", \"public_name\": \"Alex\", \"represents_as\": \"a private client\"}"

---

Privacy design

On every call the agent:

1. Introduces itself with a public alias (default: Alex) — not your name
2. Says it is calling on behalf of a private client (configurable)
3. Refuses to share your real name, phone, email, address, or workplace
4. Can still complete a task (confirm appointments, ask questions, leave messages)

Configure in .env:

AGENT_PUBLIC_NAME=Alex
AGENT_REPRESENTS_AS=a private client

---

Tuning (optional)

If short noise still cuts the agent off, raise barge-in thresholds in .env:

MIN_INTERRUPTION_DURATION=1.5
MIN_INTERRUPTION_WORDS=2
FALSE_INTERRUPTION_TIMEOUT=1.5

Other useful knobs (see .env.example):

• MIN_ENDPOINTING_DELAY / MAX_ENDPOINTING_DELAY — how fast the agent replies after you stop talking
• AUTO_HANGUP_SILENCE_SEC / AUTO_HANGUP_MAX_CALL_SEC — call length limits
• LLM_MAX_TOKENS / LLM_REASONING_EFFORT — reply length and reasoning cost
• TTS_PACE — speaking speed

---

Project layout

voice-agent/
  agent.py                 # LiveKit worker + anonymous call logic
  make_call.py             # Dispatch helper + live status
  listen.py                # Subscribe-only live listen / browser player
  call_monitor.py          # Terminal dashboard + transcript logging
  validate_voice.py        # Voice / TTS sanity checks
  patch_local_inference.py # Windows ACCESS_VIOLATION workaround
  requirements.txt
  .env.example             # Template — copy to .env locally
  .gitignore

---

Security & legal

• Never commit .env. Only ship .env.example with placeholders.
• Rotate API keys if they were ever shared in chat, screenshots, or a private fork.
• Only call numbers you are authorized to contact. Comply with local telemarketing, consent, and recording laws.

---

License

Use and modify freely for your own projects. Add a formal license file if you redistribute under specific terms.