# mknotes

A command-line tool to transcribe all MP3, M4A, and WAV audio files in a directory using Faster Whisper, then enhance the transcriptions into comprehensive notes using OpenAI's GPT-4.1 model.

## Features

- Batch transcribes all `.mp3`, `.m4a`, and `.wav` files in a specified directory
- Automatically converts WAV files to MP3 format before processing
  - Converted MP3 files are saved in the same directory as the original WAV files
  - Reuses existing MP3 files if they've already been converted
- Saves transcriptions as `.txt` files
- Enhances notes using GPT-4.1 with a custom prompt
- Outputs enhanced notes in markdown format
- Configurable input and output directories

## Installation

```bash
# Clone the repository
git clone https://github.com/yourusername/mknotes.git
cd mknotes

# Create a virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Install ffmpeg (required for WAV to MP3 conversion)
# On Ubuntu/Debian:
# sudo apt-get install ffmpeg

# On macOS with Homebrew:
# brew install ffmpeg

# On Windows:
# Download from https://ffmpeg.org/download.html and add to PATH
```

## Usage

```bash
export OPENAI_API_KEY="your-api-key-here"
python main.py --input-dir /path/to/audio/files --output-dir /path/to/output [--turbo]
```

- `--input-dir`: Directory containing audio files (.mp3, .m4a, .wav) (required)
- `--output-dir`: Directory for output files (default: "output")
- `--turbo`: Enable turbo mode for faster inference (uses int8_float16 compute type)
- `--force`: Force re-processing of files even if output files already exist

### Turbo Mode Hardware Requirements

The `--turbo` flag enables faster inference using the `int8_float16` compute type, which can significantly speed up transcription. However, this requires:

- CUDA-compatible GPU with Tensor Cores (NVIDIA Ampere, Turing, or newer architecture)
- Or CPU with AVX2 support

If your hardware does not support this optimization, the program will automatically fall back to the next most compatible compute type and print a warning.

#### Compute Type Fallback

The program will attempt to use the most efficient compute type supported by your hardware, in the following order:

- `int8_float16` (if `--turbo` is enabled)
- `float16`
- `int8`
- `float32` (most compatible, works on virtually all hardware)

If a compute type is not supported, the program will try the next one in the list until successful.

## Requirements

- Python 3.8+
- [Faster Whisper](https://github.com/SYSTRAN/faster-whisper)
- [OpenAI Python SDK](https://github.com/openai/openai-python)
- [pydub](https://github.com/jiaaro/pydub) (for WAV to MP3 conversion)
- [ffmpeg](https://ffmpeg.org/) (required by pydub for audio conversion)