CustomGPT Widget - Voice-Enabled AI Assistant

A dual-mode AI assistant with voice and chat interfaces. Interact naturally through voice with real-time particle animations or 3D avatar, or use the text-based chat interface with message reactions and citations.

Get you CustomGPT.ai RAG API key here, needed to use this integration.

Features

Voice Mode

Voice Activity Detection - Automatic speech detection, no button pressing required
3D Avatar with Lip-Sync - Realistic avatar with facial expressions and gestures (NEW!)
Particle Animation Interface - Visual feedback with dynamic particle effects
Multiple TTS Providers - Choose from OpenAI TTS, gTTS, ElevenLabs, Edge TTS, or StreamElements
Avatar States - Dynamic mood changes and gestures during conversation (listening, thinking, speaking)

Chat Mode

Text-Based Chat - Full-featured chat interface with markdown support
Message Reactions - Like/dislike AI responses for feedback
Citations - Hover over citations to see source details
Per-Message TTS - Play any message as audio with speaker button
Speech Input - Use microphone button for voice-to-text input

Shared Features

AI-Powered Responses - Powered by CustomGPT for RAG-based responses or OpenAI for general chat
Conversation Memory - Maintains context across the conversation
Multi-language Support - Configure language for both speech recognition and synthesis
Docker Ready - Easy deployment with Docker

Prerequisites

For Docker Deployment:

For Local Development:

Python 3.10+
Node.js 18+
FFmpeg (Install guide)
Yarn (npm install -g yarn)
OpenAI API key

Quick Start (Docker)

1. Create `.env` file

# Required
OPENAI_API_KEY=sk-your-key-here

# Recommended
AI_COMPLETION_MODEL=gpt-4o-mini
TTS_PROVIDER=OPENAI
OPENAI_TTS_VOICE=nova
LANGUAGE=en

2. Run container

chmod +x run.sh  # First time only
./run.sh

3. Open browser

Visit http://localhost:8000 and allow microphone access.

Docker Hub Deployment

Using Pre-built Image

The fastest way to deploy is using the pre-built Docker image:

docker pull zriyansh/customgpt-widget:latest

Supported Architectures: AMD64, ARM64 (Mac M1/M2, Raspberry Pi)

Method 1: Docker Run (Simple)

docker run -d \
  --name customgpt-widget \
  -p 8000:8000 \
  -e OPENAI_API_KEY=your_key_here \
  -e AI_COMPLETION_MODEL=gpt-4o-mini \
  -e TTS_PROVIDER=OPENAI \
  -e LANGUAGE=en \
  zriyansh/customgpt-widget:latest

Method 2: Docker Compose (Recommended)

Create docker-compose.yml:

version: '3.8'
services:
  widget:
    image: zriyansh/customgpt-widget:latest
    container_name: customgpt-widget
    ports:
      - "8000:8000"
    env_file:
      - .env
    restart: unless-stopped

Run:

docker-compose up -d

Production Deployment

With Nginx Reverse Proxy:

server {
    listen 80;
    server_name yourdomain.com;

    location / {
        proxy_pass http://localhost:8000;
        proxy_http_version 1.1;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection 'upgrade';
        proxy_set_header Host $host;
        proxy_cache_bypass $http_upgrade;
    }
}

With SSL (Let's Encrypt):

# Install certbot
sudo apt install certbot python3-certbot-nginx

# Get certificate
sudo certbot --nginx -d yourdomain.com

# Auto-renewal
sudo certbot renew --dry-run

Website Integration

Script Tag Embed (2 Steps)

Step 1: Add to your HTML (before </body>):

<script>
  window.customGPTConfig = {
    serverUrl: 'https://your-server.com',  // Your backend URL
    position: 'bottom-right',              // bottom-right, bottom-left
    theme: 'dark',                         // dark or light
    initialMode: 'chat',                   // chat or voice
    showBranding: true                     // Show "Powered by CustomGPT"
  };
</script>
<script src="https://your-server.com/widget.js" defer></script>

Step 2: Deploy your backend with Docker (see above)

Customization Options

window.customGPTConfig = {
  // Required
  serverUrl: 'https://your-server.com',

  // Optional UI
  position: 'bottom-right',       // bottom-right, bottom-left, bottom-center
  theme: 'dark',                  // dark, light
  primaryColor: '#8b5cf6',        // Brand color
  initialMode: 'chat',            // chat, voice
  showBranding: true,             // Show branding

  // Optional Behavior
  autoOpen: false,                // Auto-open on page load
  openDelay: 3000,                // Delay before auto-open (ms)
  greeting: 'Hi! How can I help?', // Initial greeting message

  // Optional Features
  enableVoiceMode: true,          // Show voice mode button
  enableSTT: true,                // Enable speech input
  enableTTS: true,                // Enable audio output
  enableAvatar: true              // Enable 3D avatar mode
};

Platform-Specific Integration

WordPress:

Install "Insert Headers and Footers" plugin
Paste script in footer section
Save and clear cache

Shopify:

Go to Online Store �� Themes �� Edit Code
Open theme.liquid
Add script before </body>
Save

Wix:

Add "Custom Code" element
Paste script in HTML iframe
Position the element

Next.js / React:

import { useEffect } from 'react';

export default function MyApp() {
  useEffect(() => {
    window.customGPTConfig = {
      serverUrl: process.env.NEXT_PUBLIC_WIDGET_URL,
      theme: 'dark'
    };

    const script = document.createElement('script');
    script.src = `${process.env.NEXT_PUBLIC_WIDGET_URL}/widget.js`;
    script.defer = true;
    document.body.appendChild(script);

    return () => document.body.removeChild(script);
  }, []);

  return <YourApp />;
}

Working Examples

For complete integration examples and step-by-step guides, see the examples/ directory:

Integration Guide - Complete documentation for website integration
Floating Widget Example - Test floating chatbot interface
Inline Embed Example - Test inline page embedding

The examples directory includes:

Platform-specific integration instructions (WordPress, Shopify, Wix, etc.)
Framework integration examples (Next.js, React, Vue)
Customization options and CSS examples
Analytics tracking setup
Troubleshooting common issues

Avatar Mode (3D Talking Avatar)

What is Avatar Mode?

Avatar Mode replaces the particle animation interface with a realistic 3D animated avatar that provides visual feedback through facial expressions, gestures, and lip-sync animation during your conversation. The avatar creates a more human-like and engaging interaction experience.

Browser Requirements

Avatar Mode requires a modern browser with WebGL support:

** Fully Supported:**

Chrome/Edge 90+ (Desktop & Mobile)
Firefox 88+ (Desktop & Mobile)
Safari 14+ (Desktop & Mobile)
Opera 76+

System Requirements:

WebGL 2.0 capable GPU
Minimum 4GB RAM recommended
Stable internet connection (avatar loads from CDN)

Check WebGL Support: Visit https://get.webgl.org/ to verify your browser supports WebGL.

Installation & Hosting

No Additional Installation Required! Avatar Mode is built-in and enabled by default.

How It Works:

3D Avatar Library: Loaded from CDN (@met4citizen/talkinghead@1.6.0)
Avatar Models: Streamed from Ready Player Me CDN
Zero Configuration: Works out-of-the-box with standard deployment
No Special Hosting: Same hosting as the rest of the application (Docker or local)

The avatar system automatically:

Detects WebGL support on startup
Loads the TalkingHead library from CDN (one-time ~200KB)
Streams the 3D avatar model (~2-3MB, cached by browser)
Falls back to particle mode if WebGL is unavailable

Customization Options

Avatar Model - Configure via environment variable:

# Optional: Custom Ready Player Me avatar URL
VITE_AVATAR_GLB_URL=https://models.readyplayer.me/YOUR_AVATAR_ID.glb

Default Avatar: High-quality female avatar with optimized morphTargets for lip-sync

Performance Tuning (frontend/src/utils/avatarConfig.ts):

Desktop: 60 FPS target
Mobile: 30 FPS target
Auto-detection of device tier (high/medium/low)
Configurable load timeout and retry settings

Technical Details

Architecture:

Library: TalkingHead v1.6.0 by @met4citizen
3D Models: Ready Player Me GLB format with ARKit/Oculus Visemes
Rendering: Three.js WebGL renderer (managed by TalkingHead)
Lip-Sync: Word-based phoneme animation with English language support
State Management: React hooks with global method exposure

CDN Dependencies:

TalkingHead: https://cdn.jsdelivr.net/npm/@met4citizen/talkinghead@1.6.0/modules/talkinghead.mjs
Avatar Model: https://models.readyplayer.me/*.glb
Three.js: Bundled with TalkingHead library

TTS Provider Options

OpenAI TTS (Recommended)

High quality, natural-sounding voices
Streaming support for low latency
Requires OpenAI API key
Multiple voice options: nova, alloy, echo, fable, onyx, shimmer

TTS_PROVIDER=OPENAI
OPENAI_TTS_MODEL=tts-1  # or tts-1-hd for higher quality
OPENAI_TTS_VOICE=nova

gTTS (Google Text-to-Speech)

Free, no API key required
Good quality, supports many languages
Slight latency due to online generation

EDGETTS (Microsoft Edge TTS)

Free, no API key required
High quality, fast, supports many voices
Good alternative to OpenAI TTS

TTS_PROVIDER=EDGETTS
EDGETTS_VOICE=en-US-EricNeural  # or en-US-JennyNeural, en-GB-SoniaNeural, etc.

ELEVENLABS

Premium, requires API key
Highest quality, most natural-sounding
Get API key at elevenlabs.io

TTS_PROVIDER=ELEVENLABS
ELEVENLABS_API_KEY=your_elevenlabs_api_key
ELEVENLABS_VOICE=EXAVITQu4vr4xnSDxMaL

STREAMELEMENTS

Free, no API key required
Basic quality
Good for testing

Local Development

Backend Setup

cd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

Backend will be available at http://localhost:8000

Frontend Setup

cd frontend
yarn install
yarn dev

Frontend will be available at http://localhost:5173

Development Commands

Backend:

uvicorn main:app --reload - Run with hot reload
python -m pytest - Run tests (if added)

Frontend:

yarn dev - Development server with hot reload
yarn build - Build for production
yarn lint - Run ESLint
yarn preview - Preview production build

AI Model Configuration

When using standard OpenAI (not CustomGPT), you can choose from several models:

Model	Best For	Speed	Context
gpt-4o-mini	Most use cases, voice assistants	Very Fast	128k
gpt-4o	High-quality responses	Fast	128k
gpt-4-turbo	Maximum quality	Medium	128k
gpt-3.5-turbo	Budget option	Very Fast	16k

Recommended for voice assistants: gpt-4o-mini (best balance of speed, quality, and cost)

Update in .env:

AI_COMPLETION_MODEL=gpt-4o-mini

Speech-to-Text Models

OpenAI's latest speech-to-text models (Released March 2025):

Model	Speed	Accuracy	Best For
gpt-4o-mini-transcribe	Very Fast	Excellent	Voice assistants (Default)
gpt-4o-transcribe	Fast	Best	Maximum accuracy, challenging audio
whisper-1	Fast	Good	Whisper v2 model

Default: gpt-4o-mini-transcribe (best balance for voice assistants)

Update in .env:

STT_MODEL=gpt-4o-mini-transcribe  # or gpt-4o-transcribe, whisper-1

CustomGPT Integration

This project supports CustomGPT for RAG-based AI responses using your custom knowledge base.

Note: OpenAI API key is always required for Whisper (speech-to-text). CustomGPT is only used for AI completions.

Using CustomGPT for AI Responses

Set environment variables in .env:

# Required for speech-to-text (Whisper)
OPENAI_API_KEY=your_openai_api_key

# Enable CustomGPT for AI responses
USE_CUSTOMGPT=true
CUSTOMGPT_PROJECT_ID=your_project_id_here
CUSTOMGPT_API_KEY=your_customgpt_api_key_here
CUSTOMGPT_STREAM=true  # Enable streaming for faster responses

The system will use:

OpenAI Whisper for speech-to-text
CustomGPT for AI completions with your RAG data
Your selected TTS provider for text-to-speech

The CustomGPT endpoint is automatically configured as: https://app.customgpt.ai/api/v1/projects/{YOUR_PROJECT_ID}

Using Standard OpenAI

To use standard OpenAI GPT models instead of CustomGPT:

USE_CUSTOMGPT=false  # or omit this line
OPENAI_API_KEY=your_openai_api_key
AI_COMPLETION_MODEL=gpt-4o-mini

Configuration

Environment Variables

Variable	Required	Default	Description
`OPENAI_API_KEY`	Yes	-	Your OpenAI API key (required for STT and TTS if using OpenAI)
`STT_MODEL`	No	`gpt-4o-mini-transcribe`	Speech-to-text model
`USE_CUSTOMGPT`	No	`false`	Enable CustomGPT for AI completions
`CUSTOMGPT_PROJECT_ID`	Conditional	-	Required if USE_CUSTOMGPT=true
`CUSTOMGPT_API_KEY`	Conditional	-	Required if USE_CUSTOMGPT=true
`CUSTOMGPT_STREAM`	No	`true`	Enable streaming for faster responses
`AI_COMPLETION_MODEL`	No	`gpt-3.5-turbo`	Model to use (only for OpenAI, not CustomGPT)
`LANGUAGE`	No	`en`	ISO-639-1 language code for STT/TTS
`TTS_PROVIDER`	No	`OPENAI`	TTS provider: OPENAI, gTTS, ELEVENLABS, STREAMELEMENTS, EDGETTS
`OPENAI_TTS_MODEL`	No	`tts-1`	OpenAI TTS model: tts-1 (fast) or tts-1-hd (quality)
`OPENAI_TTS_VOICE`	No	`nova`	OpenAI TTS voice: alloy, echo, fable, onyx, nova, shimmer
`EDGETTS_VOICE`	No	`en-US-EricNeural`	Voice for Edge TTS
`ELEVENLABS_API_KEY`	Conditional	-	Required if using ELEVENLABS
`ELEVENLABS_VOICE`	No	`EXAVITQu4vr4xnSDxMaL`	Voice ID for ElevenLabs
`VITE_UI_THEME`	No	`dark`	UI theme: dark or light
`VITE_ENABLE_VOICE_MODE`	No	`true`	Show voice mode button
`VITE_ENABLE_STT`	No	`true`	Show microphone button for STT
`VITE_ENABLE_TTS`	No	`true`	Show speaker button for TTS
`VITE_AVATAR_GLB_URL`	No	Default avatar	Custom Ready Player Me avatar GLB URL

Customizing the AI Personality

Edit the system prompt in backend/ai.py:

INITIAL_PROMPT = f"You are CustomGPT Widget - a helpful assistant with a voice interface..."

Architecture

Tech Stack

Backend:

FastAPI - Web framework
OpenAI Whisper - Speech-to-text
OpenAI GPT / CustomGPT - AI completions
Multiple TTS libraries - Text-to-speech
FFmpeg - Audio processing

Frontend:

React + TypeScript
Vite - Build tool
@ricky0123/vad-react - Voice activity detection
Canvas API - Particle animations
@met4citizen/talkinghead - 3D avatar with lip-sync
Three.js - WebGL 3D rendering (via TalkingHead)
Ready Player Me - Avatar 3D models

Troubleshooting

Microphone not working

Ensure you've granted microphone permissions in your browser
Check browser console for errors
Try HTTPS (VAD requires secure context)

No audio response

Check browser console for fetch errors
Verify backend is running: curl http://localhost:8000/
Check backend logs for TTS errors
Ensure FFmpeg is installed

Poor voice recognition

Speak clearly and at a normal pace
Reduce background noise
Minimum speech duration is 0.4 seconds
Check LANGUAGE env var matches your spoken language

Docker build fails

ARM64 Mac: Ensure Docker Desktop supports ARM
Try docker buildx build --platform linux/amd64 -t customgpt-widget .
Check disk space: df -h

API errors

Verify OPENAI_API_KEY is set correctly
Check API key has sufficient credits
For CustomGPT, verify project ID and API key

Avatar Mode issues

Avatar not loading / stuck on loading screen:

Check browser console for WebGL errors: Press F12 �� Console tab
Verify WebGL support: Visit https://get.webgl.org/
Clear browser cache and reload the page
Check network connectivity (avatar loads from CDN)
Look for [Avatar] prefixed logs in console for detailed diagnostics

Avatar shows black screen or renders incorrectly:

Update your graphics drivers to the latest version
Try a different browser (Chrome/Firefox recommended)
Disable browser extensions that might interfere with WebGL
Check GPU hardware acceleration is enabled in browser settings
Fallback: Switch to Particle mode using the mode toggle

Avatar lip-sync not working / mouth not moving:

Check browser console for [Avatar] speakAudio logs
Verify TTS is working properly (test in particle mode first)
Audio format must be compatible (MP3/WAV with proper encoding)
Check [Avatar] Audio decoded logs show valid duration and channels
Ensure network is not blocking CDN resources

Avatar gestures/moods not changing:

Look for [Avatar] setListening/setProcessing/setIdle logs in console
Verify hasSetMood and hasPlayGesture are both true in logs
Check that avatar initialization completed successfully
Try reloading the page to reinitialize the avatar
Report any [Avatar] � Failed to set errors

Performance issues / choppy animation:

Lower your browser window size (fewer pixels to render)
Close other tabs/applications to free up GPU resources
Avatar auto-adjusts to 30 FPS on mobile devices
Check [Avatar] TalkingHead instance created logs for initialization time
Consider using particle mode on low-end devices

Avatar timeout errors:

Default timeout is 10 seconds, configurable in avatarConfig.ts
Slow network may cause timeout during model download
Retry will attempt up to 2 times automatically
Check network speed: Avatar model is ~2-3MB
CDN issues: Try again later or check jsdelivr.com status

Browser compatibility check:

// Open browser console (F12) and run:
console.log('WebGL:', !!document.createElement('canvas').getContext('webgl'));
console.log('TalkingHead loaded:', !!window.TalkingHead);

Debugging tips:

Enable verbose logging: Look for [Avatar] prefixed console logs
Check initialization sequence: Library load �� Instance create �� Model load
Verify global methods: Check window.avatarSetListening exists
Test WebGL independently: Visit https://threejs.org/examples/
Report issues with full browser console logs

License

This project is licensed under the MIT License.