Local AI Dictation App

A minimal, high-performance local text-to-speech and transcription app powered by Voxtral (Mistral AI).
Built for privacy, speed, and accuracy, running completely offline on your NVIDIA GPU.

Features

• 🎙️ Real-time Dictation: Record audio directly from your microphone.
• 📁 File Upload: Support for .wav, .mp3, and other audio formats.
• ✨ AI Cleanup: Automatically removes filler words ("uhm", "ah") and corrects grammar while preserving original meaning.
• ⚡ GPU Accelerated: Runs on local NVIDIA GPUs using CUDA for lightning-fast inference.
• 🌍 Multilingual: Supports multiple languages (English, German, French, etc.) by detecting the original audio language.
• 🔊 Read Aloud: Built-in Text-to-Speech to read back your transcriptions.
• 🌓 Modern UI: Clean, dark-mode interface for distraction-free writing.

Tech Stack

• Frontend: HTML5, CSS3, Vanilla JavaScript (No heavy frameworks)
• Backend: Python FastAPI
• AI Model: mistralai/Voxtral-Mini-3B-2507
• Inference: PyTorch (CUDA), Transformers

Prerequisites

• OS: Windows (Tested), Linux/Mac (Adaptations required)
• GPU: NVIDIA GPU with CUDA support (Recommended for speed).
• Python: 3.10+
• Model Files:
  • The app looks for model files in backend/model/.
  • You can download them from Hugging Face.
  • If empty, it will verify your Hugging Face credentials and download automatically (requires access approval).

Installation

1. Clone/Download this repository.

2. Setup: Run the included batch script to create a virtual environment and install dependencies.

   Double-click run_app.bat

   Note: On first run, it may download ~2.5GB of PyTorch files to enable GPU support.

Usage

1. Run run_app.bat.
2. The browser will open automatically at http://localhost:8000.
3. Record: Click the microphone icon to start/stop recording.
4. Transcribe: The app will process the audio locally and display the cleaned text.
5. Edit/Copy: Use the text area to make final distincts or copy to clipboard.

Customization

• System Prompt: Edit backend/main.py to change how the AI cleans up text (e.g., make it summarize instead of transcribe).
• Model: Change MODEL_ID in backend/main.py to use a different Voxtral size (e.g., Voxtral-Base).

Troubleshooting

• "CUDA is NOT available": Ensure you have up-to-date NVIDIA drivers. The run_app.bat script enforces specific CUDA 12.1 compatible PyTorch installation.
• Slow Transcription: Transcription on CPU is significantly slower. GPU is highly recommended.