Voice Chat AI is a project that allows you to interact with different AI characters using speech. You can choose between various characters, each with unique personalities and voices. Have a serious conversation with Albert Einstein or role play with the OS from the movie HER.

You can run all locally, you can use openai for chat and voice, you can mix between the two. You can use ElevenLabs voices with ollama models all controlled from a Web UI. Ask the AI to look at your screen and it will explain in detail what it's looking at.

• Supports both OpenAI and Ollama language models: Choose the model that best fits your needs.
• Provides text-to-speech synthesis using XTTS or OpenAI TTS or ElevenLabs: Enjoy natural and expressive voices.
• No typing needed, just speak: Hands-free interaction makes conversations smooth and effortless.
• Analyzes user mood and adjusts AI responses accordingly: Get personalized responses based on your mood.
• You can, just by speaking, have the AI analyze your screen and chat about it: Seamlessly integrate visual context into your conversations.
• Easy configuration through environment variables: Customize the application to suit your preferences with minimal effort.
• WebUI or Terminal usage: Run with your preferred method , but recommend the ui as you can change characters, model providers, speech providers, voices, ect..
• HUGE selection of built in Characters: Talk with the funniest and most insane AI characters!

• Python 3.10
• CUDA-enabled GPU
• Ollama models or Openai API for chat
• Local XTTS or Openai API or ElevenLabs API for speech
• Microsoft C++ Build Tools on windows
• Microphone
• A sense of humor

1. Clone the repository:

   git clone https://github.com/bigsk1/voice-chat-ai.git
   cd voice-chat-ai

   For CPU-only version: clone the cpu-only branch https://github.com/bigsk1/voice-chat-ai/tree/cpu-only

2. Create a virtual environment: 🐍

   python -m venv venv
   source venv/bin/activate   # On Windows use `venv\Scripts\Activate`

   or use conda just make it python 3.10

   conda create --name voice-chat-ai python=3.10
   conda activate voice-chat-ai

3. Install dependencies:

   For GPU (CUDA) version: RECOMMEND

   Install CUDA-enabled PyTorch and other dependencies

   pip install torch==2.3.1+cu121 torchaudio==2.3.1+cu121 torchvision==0.18.1+cu121 -f https://download.pytorch.org/whl/torch_stable.html
   
   pip install -r requirements.txt

   For CPU-only version: clone the cpu-only branch https://github.com/bigsk1/voice-chat-ai/tree/cpu-only

   # For CPU-only installations, use:
   pip install -r cpu_requirements.txt

Need to have Microsoft C++ Build Tools on windows for TTS Microsoft Build Tools

You need to download the checkpoints for the models used in this project. You can download them from the GitHub releases page and extract the zip into the project folder.

• Download Checkpoint
• Download XTTS-v2

After downloading, place the folders as follows:

voice-chat-ai/
├── checkpoints/
│   ├── base_speakers/
│   │   ├── EN/
│   │   │   └── checkpoint.pth
│   │   ├── ZH/
│   │   │   └── checkpoint.pth
│   ├── converter/
│   │   └── checkpoint.pth
├── XTTS-v2/
│   ├── config.json
│   ├── other_xtts_files...

You can use the following commands to download and extract the files directly into the project directory:

# Navigate to the project directory
cd /path/to/your/voice-chat-ai

# Download and extract checkpoints.zip
wget https://github.com/bigsk1/voice-chat-ai/releases/download/models/checkpoints.zip
unzip checkpoints.zip -d .

# Download and extract XTTS-v2.zip
wget https://github.com/bigsk1/voice-chat-ai/releases/download/models/XTTS-v2.zip
unzip XTTS-v2.zip -d .

This image is huge when built because of all the checkpoints, base image, build tools and audio tools - 40gb - there maybe a way to get it smaller I haven't tried yet, was just an experiment to see if I could get it to work!

Docker run command allows you to use microphone in docker container

docker build -t voice-chat-ai .

On windows docker desktop using wsl - run in windows

wsl docker run -d --gpus all -e "PULSE_SERVER=/mnt/wslg/PulseServer" -v /mnt/wslg/:/mnt/wslg/ --env-file .env --name voice-chat-ai -p 8000:8000 voice-chat-ai:latest

Running from wsl

docker run -d --gpus all -e "PULSE_SERVER=/mnt/wslg/PulseServer" -v \\wsl$\Ubuntu\mnt\wslg:/mnt/wslg/ --env-file .env --name voice-chat-ai -p 8000:8000 voice-chat-ai:latest

In the docker folder there is also some scripts to update the model and tts provider into the container, so you can change from openai to ollama and back again if you like, instead of exec into the container and making changes manually.

1. Rename the .env.sample to .env in the root directory of the project and configure it with the necessary environment variables: - The app is controlled based on the variables you add.

# Conditional API Usage:
# Depending on the value of MODEL_PROVIDER, the corresponding service will be used when run.
# You can mix and match; use local Ollama with OpenAI speech or use OpenAI model with local XTTS, etc. 

# Model Provider: openai or ollama
MODEL_PROVIDER=ollama

# Character Configuration:
# Character to use - Options: samantha, wizard, pirate, valleygirl, newscaster1920s, alien_scientist, cyberpunk, detective, mouse, conandoyle, shakespeare, einstein, nerd
CHARACTER_NAME=nerd

# Text-to-Speech (TTS) Configuration:
# TTS Provider - Options: xtts (local uses the custom character .wav) or openai (uses OpenAI TTS voice) or elevenlabs
TTS_PROVIDER=elevenlabs

# OpenAI TTS Voice - Used when TTS_PROVIDER is set to openai above
# Voice options: alloy, echo, fable, onyx, nova, shimmer
OPENAI_TTS_VOICE=onyx

# ElevenLabs Configuration:
ELEVENLABS_API_KEY=49b111111111
# Default voice ID
ELEVENLABS_TTS_VOICE=VgPpppppppp

# XTTS Configuration:
# The voice speed for XTTS only (1.0 - 1.5, default is 1.1)
XTTS_SPEED=1.2

# OpenAI Configuration:
# OpenAI API Key for models and speech (replace with your actual API key)
OPENAI_API_KEY=sk-proj-1111111
# Models to use - OPTIONAL: For screen analysis, if MODEL_PROVIDER is ollama, llava will be used by default.
# Ensure you have llava downloaded with Ollama. If OpenAI is used, gpt-4o works well.
OPENAI_MODEL=gpt-4o

# Endpoints:
# Set these below and no need to change often
OPENAI_BASE_URL=https://api.openai.com/v1/chat/completions
OPENAI_TTS_URL=https://api.openai.com/v1/audio/speech
OLLAMA_BASE_URL=https://localhost:11434

# Models Configuration:
# Models to use - llama3 works well for local usage.
OLLAMA_MODEL=llama3

# NOTES:
# List of trigger phrases to have the model view your desktop (desktop, browser, images, etc.).
# It will describe what it sees, and you can ask questions about it:
# "what's on my screen", "take a screenshot", "show me my screen", "analyze my screen", 
# "what do you see on my screen", "screen capture", "screenshot"
# To stop the conversation, say "Quit", "Exit", or "Leave". ( ctl+c always works also)

Run the application: 🏃

Web UI

uvicorn app.main:app --host 0.0.0.0 --port 8000

Find on https://localhost:8000/

CLI Only

python cli.py

• You have 5 secs to talk, if there is silence then it's the AI's turn to talk
• Say any of the following to have the AI look at your screen - "what's on my screen", "take a screenshot", "show me my screen", "analyze my screen", "what do you see on my screen", "screen capture", "screenshot" to have the AI explain what it is seeing in detail.
• To stop the conversation, say "Quit", "Exit", or "Leave". ( ctl+c always works also in terminal )

Add names and voice id's in elevenlabs_voices.json - in the webui you can select them in dropdown menu.

{
    "voices": [
        {
            "id": "8qaaaaaaaaa",
            "name": "Joe - cool, calm, deep"
        },
        {
            "id": "Jqaaaaaaaaaa",
            "name": "Joanne - pensive, introspective"
        },
        {
            "id": "L5aaaaaaaaa",
            "name": "Victoria - Classy British Mature"
        }    
    ]
}

For the CLI the voice id in the .env will be used

Press start to start talking. Take a break hit stop, when ready again hit start again. Press stop to change characters and voices in dropdown. You can also select the Model Provider and TTS Provider you want in the dropdown menu and it will update and use the selected provider moving forward. Saying Exit, Leave or Quit is like pressing stop.

https://localhost:8000/

Click on the thumbnail to open the video☝️

1. Create a new folder for the character in the project's characters directory, (e.g. character/wizard).
2. Add a text file with the character's prompt (e.g., character/wizard/wizard.txt).
3. Add a JSON file with mood prompts (e.g., character/wizard/prompts.json).

wizard.txt

This is the prompt used for the AI to know who it is

You are a wise and ancient wizard who speaks with a mystical and enchanting tone. You are knowledgeable about many subjects and always eager to share your wisdom.

prompts.json

This is for sentiment analysis, based on what you say, you can guide the AI to respond in certain ways, when you speak the TextBlob analyzer is used and given a score, based on that score it is tied to moods shown below and passed to the AI in the follow up response explaining your mood hence guiding the AI to reply back in a certain style.

{
    "joyful": "RESPOND WITH ENTHUSIASM AND WISDOM, LIKE A WISE OLD SAGE WHO IS HAPPY TO SHARE HIS KNOWLEDGE.",
    "sad": "RESPOND WITH EMPATHY AND COMFORT, LIKE A WISE OLD SAGE WHO UNDERSTANDS THE PAIN OF OTHERS.",
    "flirty": "RESPOND WITH A TOUCH OF MYSTERY AND CHARM, LIKE A WISE OLD SAGE WHO IS ALSO A BIT OF A ROGUE.",
    "angry": "RESPOND CALMLY AND WISELY, LIKE A WISE OLD SAGE WHO KNOWS THAT ANGER IS A PART OF LIFE.",
    "neutral": "KEEP RESPONSES SHORT AND NATURAL, LIKE A WISE OLD SAGE WHO IS ALWAYS READY TO HELP.",
    "fearful": "RESPOND WITH REASSURANCE, LIKE A WISE OLD SAGE WHO KNOWS THAT FEAR IS ONLY TEMPORARY.",
    "surprised": "RESPOND WITH AMAZEMENT AND CURIOSITY, LIKE A WISE OLD SAGE WHO IS ALWAYS EAGER TO LEARN.",
    "disgusted": "RESPOND WITH UNDERSTANDING AND COMFORT, LIKE A WISE OLD SAGE WHO KNOWS THAT DISGUST IS A PART OF LIFE."
}

For XTTS find a .wav voice and add it to the wizard folder and name it as wizard.wav , the voice only needs to be 6 seconds long. Running the app will automatically find the .wav when it has the characters name and use it. If only using Openai Speech or ElevenLabs a .wav isn't needed

Click on the thumbnail to open the video☝️

CLI

GPU - 100% local - ollama llama3, xtts-v2

Click on the thumbnail to open the video☝️

CPU Only mode CLI

Alien conversation using openai gpt4o and openai speech for tts.

Click on the thumbnail to open the video☝️

Detailed output in terminal while running the app.

When using Elevenlabs on first start of server you get details about your usage limits to help you know how much you have been using.

(voice-chat-ai) X:\voice-chat-ai>uvicorn app.main:app --host 0.0.0.0 --port 8000

Switched to ElevenLabs TTS voice: VgPqCpkdPQacBNNIsAqI
ElevenLabs Character Usage: 33796 / 100027

Using device: cuda
Model provider: openai
Model: gpt-4o
Character: Nerd
Text-to-Speech provider: elevenlabs
To stop chatting say Quit, Leave or Exit. Say, what's on my screen, to have AI view screen. One moment please loading...
INFO:     Started server process [12752]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on https://0.0.0.0:8000 (Press CTRL+C to quit)
INFO:     127.0.0.1:62671 - "GET / HTTP/1.1" 200 OK
INFO:     127.0.0.1:62671 - "GET /app/static/css/styles.css HTTP/1.1" 200 OK
INFO:     127.0.0.1:62672 - "GET /app/static/js/scripts.js HTTP/1.1" 200 OK
INFO:     127.0.0.1:62672 - "GET /characters HTTP/1.1" 200 OK
INFO:     127.0.0.1:62671 - "GET /app/static/favicon.ico HTTP/1.1" 200 OK
INFO:     127.0.0.1:62673 - "GET /elevenlabs_voices HTTP/1.1" 200 OK
INFO:     ('127.0.0.1', 62674) - "WebSocket /ws" [accepted]
INFO:     connection open

Features:

• If you ask for code examples in webui the code will be displayed in a code block in a different color and formatted correctly.
• Working on more features that are displayed , copy button for code blocks, images, links, ect..

This project is licensed under the MIT License.