🎙 OpenVoice AI

A real-time, multi-agent voice assistant powered by LangGraph Swarm architecture. Speak naturally to specialized AI agents — each with their own personality, voice, and expertise — through a stunning animated web interface.

🌟 Features

Feature	Description
🗣️ Real-time Voice	Speak and listen in real-time via the browser using WebRTC — VAD detects speech, ASR transcribes, LLM responds, TTS speaks back
🤖 Multi-Agent Swarm	3 specialized agents (Customer Care, Shopper, Order Ops) with seamless handoffs via LangGraph
🎨 Animated Agent Orbs	Each agent has a unique color identity with animated orb visualizations that glow when speaking
💬 Dual Mode	Switch between Voice mode (mic) and Text mode (typing) on the fly
🔄 Live Agent Handoffs	Agents transfer conversations to each other based on context — orb morphs colors during handoff
🧠 Conversation Memory	In-memory checkpointer (`MemorySaver`) persists conversation state across the session for zero latency
⚡ GPU Accelerated	VAD (Silero), ASR (Qwen3-0.6B), and TTS (Piper) all run on CUDA when available
🎤 Interrupt Support	Interrupt the AI mid-sentence — it remembers where it was cut off
🔇 Noise Reduction	Real-time audio denoising using DeepFilterNet3 ONNX model (toggleable via UI)
🌐 Language Guard	Production-grade English language enforcement gate using `langid` to drop spurious noise/translations
⏱️ Latency Tracking	Granular latency metrics tracking (TTFT, TTFA, and agent switching)

🏗️ Architecture

graph TB
    subgraph Frontend["React Frontend (Vite)"]
        UI[Animated UI Orbs]
        Audio[Web Audio API]
        RTC_Client[WebRTC Client]
        WS_Client["WebSocket Client (Text Chat)"]
        
        Audio <--> RTC_Client
        RTC_Client <--> UI
        WS_Client <--> UI
    end

    subgraph Backend["FastAPI Backend"]
        WS_Server["WebSocket Endpoint"]
        WebRTC["WebRTC Endpoint"]
        Pipeline[WebVoicePipeline]
        
        subgraph Models["ML Models"]
            VAD["Silero VAD"]
            ASR["Qwen3 ASR"]
            TTS["Piper TTS"]
        end
        
        subgraph Agents["LangGraph Swarm"]
            State[("MemorySaver Checkpointer")]
            CC["CustomerCare (Alpha)"]
            Shop["Shopper (Gamma)"]
            Ops["OrderOps (Beta)"]
            
            CC <--> State
            Shop <--> State
            Ops <--> State
        end
        
        WS_Server <--> Pipeline
        WebRTC <--> Pipeline
        Pipeline <--> Models
        Pipeline <--> Agents
    end

    WS_Client <-->|ws://chat| WS_Server
    RTC_Client <-->|WebRTC Data Channels| WebRTC

🤖 The Agents

Agent	Voice	Color	Specialization	Tools
Customer Care (Alpha)	🇬🇧 Alba (British)	🟣 `#4B8DFF`	Returns, refunds, policies, general help	`lookup_policy`, transfer tools
Shopper (Gamma)	🇺🇸 Bryce (American)	🟢 `#00C9A7`	Product search, recommendations, catalog	`search_catalog`, transfer tools
Order Ops (Beta)	🇺🇸 HFC Female	🔴 `#FF6FAE`	Order tracking, delivery status, operations	`check_order_status`, transfer tools

Each agent can transfer seamlessly to another via LangGraph tool calls. The user never notices the handoff — the orb simply morphs its color.

📂 Project Structure

OpenVoice AI/
├── backend/
│   ├── src/
│   │   ├── api/                    # 🌐 FastAPI + WebSocket layer
│   │   │   ├── server.py           # FastAPI app, WS endpoints, REST routes
│   │   │   └── web_pipeline.py     # WebSocket-adapted voice pipeline
│   │   ├── agents/                 # 🤖 LangGraph agent system
│   │   │   ├── session.py          # VoiceSession — LangGraph graph builder
│   │   │   ├── state.py            # VoiceState TypedDict
│   │   │   └── specialized/        # Individual agent definitions
│   │   │       ├── customer_care.py
│   │   │       ├── shopper.py
│   │   │       └── order_ops.py
│   │   ├── asr/                    # 🎤 Automatic Speech Recognition
│   │   │   └── whisper.py          # Qwen3-ASR-0.6B model wrapper
│   │   ├── audio/                  # 🔊 Audio processing & I/O
│   │   │   ├── io.py               # sounddevice mic/speaker (CLI only)
│   │   │   └── denoiser.py         # DeepFilterNet3 ONNX audio denoiser
│   │   ├── core/                   # ⚙️ Core abstractions
│   │   │   ├── interfaces.py       # IVAD, IASR, ILLM, ITTS interfaces
│   │   │   └── pipeline.py         # Original CLI voice pipeline
│   │   ├── llm/                    # 🧠 LLM client
│   │   │   └── client.py           # LLMModel — wraps VoiceSession
│   │   ├── tts/                    # 🗣️ Text-to-Speech
│   │   │   └── piper.py            # Piper TTS (ONNX, GPU-accelerated)
│   │   ├── utils/                  # 🛠️ Utilities
│   │   │   ├── chunker.py          # SentenceChunker for TTS streaming
│   │   │   └── language_guard.py   # English language enforcement gate
│   │   └── vad/                    # 🎯 Voice Activity Detection
│   │       └── silero.py           # Silero VAD (PyTorch, GPU)
│   ├── models/                     # 📦 Downloaded TTS voice models
│   └── .env                        # API keys (not committed)
│
├── frontend/
│   ├── src/
│   │   ├── components/             # ⚛️ React components
│   │   │   ├── VoiceOrb.jsx        # Animated agent orb + particles
│   │   │   ├── AgentLabel.jsx      # Agent name + status badge
│   │   │   ├── MicButton.jsx       # Mic toggle with pulse animation
│   │   │   ├── TranscriptPanel.jsx # Conversation sidebar
│   │   │   ├── TextInputBar.jsx    # Text chat input
│   │   │   ├── ConnectionStatus.jsx# WebSocket status dot
│   │   │   ├── NoiseReductionToggle.jsx # UI toggle for audio denoiser
│   │   │   └── ModeToggle.jsx      # Voice ↔ Text switch
│   │   ├── hooks/                  # 🪝 Custom React hooks
│   │   │   ├── useWebSocket.js     # WebSocket connection management (Text)
│   │   │   ├── useWebRTC.js        # WebRTC connection management (Voice)
│   │   │   ├── useAudio.js         # Mic capture + TTS playback
│   │   │   └── useVoicePipeline.js # Orchestration hook
│   │   ├── config/
│   │   │   └── agents.js           # Agent metadata constants
│   │   ├── App.jsx                 # Root component
│   │   ├── main.jsx                # React entry point
│   │   └── index.css               # Full design system + animations
│   ├── index.html
│   ├── vite.config.js
│   └── package.json
│
├── pyproject.toml                  # Python dependencies (uv)
├── uv.lock                         # Locked dependency versions
│
└── .gitignore

🚀 Getting Started

Prerequisites

Tool	Version	Purpose
Python	3.12.x	Backend runtime
uv	Latest	Python package manager
Node.js	18+	Frontend tooling
CUDA	11.8+	GPU acceleration (optional)

1. Clone the Repository

git clone https://github.com/BadrinathanTV/OpenVoice-AI.git
cd "OpenVoice AI"

2. Backend Setup

cd backend

# Create .env from the example
cp .env.example .env

# Edit .env with your API keys
nano .env

Required .env values:

OPENAI_API_KEY=sk-your-openai-key
GROQ_API_KEY=gsk_your-groq-key
ASR_MODEL_PATH=/path/to/Qwen3-ASR-0.6B
ASR_BACKEND=transformers
ASR_STREAMING_CHUNK_SIZE_SEC=0.64
DATABASE_URL=mongodb://localhost:27017/

Install dependencies and start with uv only:

# Run this from the repository root
uv sync

# Then start the backend from backend/
cd backend
uv run --project .. uvicorn src.api.server:app --host 0.0.0.0 --port 8000

.python-version pins the repo to Python 3.12, so cloud machines and client machines should use that same interpreter line for the most reliable install.

The first run will:

Download Qwen3-ASR-0.6B model
Download Piper TTS voice models (~50MB each)
Install all Python dependencies

To enable Qwen streaming ASR, switch to the vLLM backend:

ASR_BACKEND=vllm
ASR_STREAMING_CHUNK_SIZE_SEC=0.64

Install the streaming stack with uv before starting the backend:

uv sync --extra streaming-asr

The streaming-asr extra is intended for Linux GPU environments, which matches the current CUDA-based deployment path for this project.

3. Frontend Setup

cd frontend

# Install Node.js dependencies
npm install

# Start development server
npm run dev

4. Open in Browser

Navigate to http://localhost:5173 — you'll see the animated orb UI.

Click the 🎤 mic button to start talking
Or switch to 💬 Text mode to type messages
Watch the orb glow and pulse when the AI speaks

🎨 Frontend Animation States

The agent orb transitions through visual states:

State	Animation	When
Idle	Gentle breathing pulse	Waiting for user input
Listening	Concentric ring ripples	Mic active, capturing audio
Processing	Spinning orbital rings	ASR transcribing speech
Thinking	Color desaturation + spin	Waiting for LLM response
Speaking	Full glow burst + particles	TTS audio playing back
Handoff	Color morph crossfade	Agent transferring to another

🔌 API Endpoints

REST

Method	Path	Description
`GET`	`/api/health`	Health check, shows if models are loaded
`GET`	`/api/agents`	Returns list of available agents with metadata

Realtime Transport (WebRTC / WebSocket)

Path	Mode	Protocol
`/api/webrtc/offer`	Voice mode (primary)	WebRTC Data Channels (Binary PCM + JSON)
`/ws/voice`	Voice mode (fallback)	WebSocket (Binary PCM + JSON)
`/ws/chat`	Text mode	WebSocket (JSON only)

Control message types (server → client):

{"type": "session", "threadId": "...", "agent": "CustomerCare"}
{"type": "status", "value": "recording|processing|thinking|speaking|idle"}
{"type": "agent", "name": "Shopper"}
{"type": "transcript", "role": "user|ai", "text": "...", "agent": "...", "partial": true|false}
{"type": "audio", "data": "<base64 PCM>", "sampleRate": 22050}

🛠️ Tech Stack

Backend

Component	Technology
Runtime	Python 3.11+
Web Server	FastAPI + Uvicorn
WebRTC	aiortc
Agent Framework	LangGraph Swarm
LLM	OpenAI GPT-4o-mini
ASR	Qwen3-ASR-0.6B (GPU)
TTS	Piper TTS (ONNX, GPU)
VAD	Silero VAD (PyTorch, GPU)
Denoising	DeepFilterNet3 (ONNX)
Language ID	langid
Database	In-memory (MemorySaver)
Package Manager	uv

Frontend

Component	Technology
Framework	React 19 + Vite
Styling	Vanilla CSS (design tokens)
Audio	Web Audio API
Communication	WebRTC (Data Channels) / WebSocket
Font	Inter (Google Fonts)

🧱 SOLID Principles

The codebase follows SOLID design principles:

Single Responsibility — Each component, hook, and module has one job (e.g., VoiceOrb only renders, useAudio only handles audio)
Open/Closed — Agent config in agents.js is extendable without modifying components. Add a new agent by adding an entry.
Liskov Substitution — All backend modules implement abstract interfaces (IVAD, IASR, ILLM, ITTS). Swap implementations freely.
Interface Segregation — useVoicePipeline exposes a clean API without leaking WebRTC/WebSocket or Audio internals to components.
Dependency Inversion — React components receive data via props from hooks, not from globals. Backend pipeline depends on interfaces, not concrete classes.

📄 Voice Pipeline Flow

flowchart TD
    subgraph Browser["Browser Client"]
        Mic["Microphone API"]
        Speak["Audio Playback"]
        Mic -->|32ms PCM chunks| RTC["WebRTC Data Channels"]
        RTC -->|Base64 PCM| Speak
    end

    subgraph Backend["FastAPI Backend"]
        RTC --> VAD{Silero VAD}
        VAD -->|Noise| Drop[Discard]
        VAD -->|"speech (vol > 0.005)"| Buf[Audio Buffer]
        
        Buf -->|Complete Phrase| ASR["Qwen3 ASR"]
        ASR -->|Text| Swarm{LangGraph Swarm}
        
        Swarm -->|Token Stream| Chunker[Sentence Chunker]
        Chunker -->|Complete Sentences| TTS["Piper TTS"]
        TTS -->|Audio Bytes| RTC
    end
    
    subgraph Agents["LangGraph Agents"]
        Swarm <--> CC["Customer Care (Alpha)"]
        Swarm <--> SH["Shopper (Gamma)"]
        Swarm <--> OO["Order Ops (Beta)"]
    end

🔧 Configuration

Environment Variables

Variable	Required	Default	Description
`OPENAI_API_KEY`	✅	—	OpenAI API key for GPT-4o-mini
`GROQ_API_KEY`	❌	—	Groq API key (alternative LLM provider)
`ASR_MODEL_PATH`	❌	`Qwen3-ASR-0.6B`	Path to local ASR model
`ASR_BACKEND`	❌	`transformers`	ASR backend: `transformers` or `vllm`
`ASR_STREAMING_CHUNK_SIZE_SEC`	❌	`0.64`	Streaming ASR decode chunk size in seconds
`ASR_ALLOW_BACKEND_FALLBACK`	❌	`true`	Fall back to transformers if vLLM ASR fails to initialize
`ASR_ENFORCE_ENGLISH`	❌	`true`	Enable backend language enforcement gate
`ASR_ENGLISH_CONFIDENCE_THRESHOLD`	❌	`0.80`	Language ID confidence threshold
`DATABASE_URL`	❌	`mongodb://localhost:27017/`	MongoDB connection URL

Adding a New Agent

Create backend/src/agents/specialized/your_agent.py with a system prompt, tools, and get_your_agent() function
Register it in backend/src/agents/session.py (add node + routing)
Add TTS voice in backend/src/api/web_pipeline.py
Add agent metadata in frontend/src/config/agents.js

📜 License

This project is for educational and research purposes.

Built with ❤️ by The Three !

Name		Name	Last commit message	Last commit date
Latest commit History 31 Commits
.vscode		.vscode
backend		backend
frontend		frontend
.gitignore		.gitignore
.python-version		.python-version
Dockerfile		Dockerfile
README.md		README.md
benchmarks.jpeg		benchmarks.jpeg
entrypoint.sh		entrypoint.sh
pyproject.toml		pyproject.toml
pyrightconfig.json		pyrightconfig.json
start.bat		start.bat
start.sh		start.sh
uv.lock		uv.lock

Navigation Menu

Repository files navigation

🎙 OpenVoice AI

🌟 Features

🏗️ Architecture

🤖 The Agents

📂 Project Structure

🚀 Getting Started

Prerequisites

1. Clone the Repository

2. Backend Setup

3. Frontend Setup

4. Open in Browser

🎨 Frontend Animation States

🔌 API Endpoints

REST

Realtime Transport (WebRTC / WebSocket)

🛠️ Tech Stack

Backend

Frontend

🧱 SOLID Principles

📄 Voice Pipeline Flow

🔧 Configuration

Environment Variables

Adding a New Agent

📜 License

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

🎙 OpenVoice AI

🌟 Features

🏗️ Architecture

🤖 The Agents

📂 Project Structure

🚀 Getting Started

Prerequisites

1. Clone the Repository

2. Backend Setup

3. Frontend Setup

4. Open in Browser

🎨 Frontend Animation States

🔌 API Endpoints

REST

Realtime Transport (WebRTC / WebSocket)

🛠️ Tech Stack

Backend

Frontend

🧱 SOLID Principles

📄 Voice Pipeline Flow

🔧 Configuration

Environment Variables

Adding a New Agent

📜 License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages