| README.md |
|
|
| 🛡️ Police Bot Runtime — AI Voice Assistant for Bengaluru Police |
| This is the runtime layer of an AI-powered mental wellness assistant built for frontline Bengaluru Police officers. The assistant runs locally and privately using an LLM (via Ollama) and a fine-tuned voice cloning TTS model (via XTTS/F5-TTS), producing empathetic voice replies in real-time based on officer input. |
|
|
| 💬 Text In → 🤖 LLM Reply → 🗣️ Voice Cloned Output |
|
|
| 📌 Project Goals |
| Create a voice-first wellness chatbot for police personnel |
|
|
| Fully private, runs entirely offline on powerful local machines |
|
|
| Replies are generated by an open-source LLM (police-bot) via Ollama |
|
|
| Responses are spoken aloud using a fine-tuned XTTS voice cloned from a real speaker |
|
|
| Easily extendable to integrate into a React-based web interface later |
|
|
| Future support for Kannada via multi-lingual XTTS fine-tuning |
|
|
| 🧠 System Overview |
| This repo powers the runtime experience. |
|
|
| Ollama runs an LLM (LLaMA 3, Mistral etc.) via the police-bot model |
|
|
| Python script police_runtime.py communicates with Ollama (localhost:11434) |
| |
| LLM reply is sent to police_bot_voice.py |
| |
| XTTS reads a voice reference and generates a realistic audio response (output.wav) |
| |
| Audio is played back to the officer |
| |
| Example flow: |
| |
| Officer: I'm feeling low today |
| Assistant: [spoken aloud] Namaskara! I’m here to support you. You're a valued member of the force... |
| |
| 📂 Folder Structure |
| Your folder layout should look like: |
| |
| police-bot-runtime/ |
| │ |
| ├── police_runtime.py # Main loop: user input → LLM → voice |
| ├── police_bot_voice.py # Loads XTTS model, speaks response |
| │ |
| ├── my_finetuned_model/ # XTTS fine-tuned model files |
| │ ├── config.json |
| │ ├── dvae.ptj |
| │ ├── mel_stats.pth |
| │ ├── model.pth |
| │ └── vocab.json |
| │ |
| ├── datasets-1/ |
| │ └── wavs/ |
| │ └── 0029.wav # Reference voice clip used for inference |
| │ |
| ├── venv/ # Python virtual environment |
| └── requirements.txt # (Optional) Dependency list |
| |
| ⚙️ Setup Instructions |
| 🧩 Prerequisites: |
| |
| Windows 10/11, 64-bit |
| |
| Python 3.11 (recommended) |
| |
| Ollama installed: https://ollama.com |
| |
| Trained XTTS voice model (via F5-TTS or Coqui) |
| |
| Clone or copy this folder as police-bot-runtime |
| |
| Create a virtual environment: |
| |
| bash |
| Copy |
| Edit |
| python -m venv venv |
| venv\Scripts\activate |
| Install dependencies: |
| |
| bash |
| Copy |
| Edit |
| pip install TTS requests |
| Start the Ollama LLM (separate terminal): |
| |
| bash |
| Copy |
| Edit |
| ollama run police-bot |
| Run the voice assistant: |
| |
| bash |
| Copy |
| Edit |
| python police_runtime.py |
| You'll be prompted for input. Replies will be printed and voiced out. |
|
|
| 🧪 Testing Voice Model |
| Your test_voice.py should look like: |
| |
| python |
| Copy |
| Edit |
| from TTS.api import TTS |
| |
| tts = TTS( |
| model_path="my_finetuned_model/", |
| config_path="my_finetuned_model/config.json", |
| gpu=True |
| ) |
| |
| tts.tts_to_file( |
| text="Hey officer, everything okay?", |
| file_path="output.wav", |
| speaker_wav="datasets-1/wavs/0029.wav", |
| language="en" |
| ) |
| 🧠 How It Works |
| Ollama LLM (LLaMA3) is launched with ollama run police-bot |
| |
| Text prompt is sent to http://localhost:11434/api/generate |
|
|
| Response is sent to XTTS model for synthesis |
|
|
| XTTS clones voice using speaker_wav file |
| |
| Audio saved to output.wav and played using playsound or equivalent |
| |
| 🛠️ Future Enhancements |
| Integrate directly with the frontend (React + ShadCN + Web Speech API) |
| |
| Support both English and Kannada voice cloning |
| |
| Add emotional sentiment analysis (LLaMA + classifier) |
| |
| Handle stress journaling and daily wellness logging |
| |
| Auto-log sessions to Supabase or local DB |
| |
| Deploy offline on a rugged laptop for field use |
| |
| 🙏 Acknowledgements |
| Voice cloning: F5-TTS / Coqui XTTS |
| |
| LLM inference: Ollama + LLaMA3 8B |
| |
| Frontend inspiration: GovSchemes project |
| |
| This is a labor of respect and service. Designed to support those who protect us. |
| |
| — |
