wave2midi

wave2midi is a Python tool that converts WAV audio files to MIDI format. It leverages a state-of-the-art source separation model, Demucs, to first separate the audio into individual instrument tracks (vocals, drums, bass, and other). This separation allows for more accurate pitch detection and results in cleaner, more usable MIDI files for each instrument.

Features

• High-Quality Source Separation: Uses Demucs to separate input WAV files into four distinct stems: vocals, drums, bass, and other.
• Accurate Note Detection: Converts each isolated stem to a corresponding MIDI file with accurate note detection using librosa.
• Preserves Musical Information: Retains timing and note velocity information in the output MIDI files.
• Configurable: Allows for configuration of MIDI conversion parameters via a JSON file.

How It Works

1. Stem Separation: The input WAV file is processed by Demucs, which isolates the vocals, drums, bass, and other instrument tracks.
2. Audio Analysis: Each stem undergoes pitch detection (librosa.pyin) to identify musical notes.
3. MIDI Conversion: The detected notes are converted to MIDI events with appropriate timing, velocity, and instrument information using the mido library.
4. File Output: Each stem produces a separate MIDI file, allowing for individual editing and remixing. For an input file song.wav, the tool will produce song_vocals.mid, song_drums.mid, song_bass.mid, and song_other.mid.

Dependencies

• Python 3.8+
• Demucs (and its dependencies, including torch and torchaudio)
• Librosa
• Mido
• NumPy

Installation

1. Clone the repository:

   git clone https://github.com/your-username/wave2midi.git
   cd wave2midi

2. Install the package: The project is packaged with setup.py. You can install it using pip.

   For users:

   pip install .

   This will install the wave2midi command and its dependencies. Note that this may take some time as it needs to download the Demucs models and PyTorch.

   For developers: If you want to install the project in editable mode and get the test dependencies, use the requirements.txt file:

   pip install -r requirements.txt

Usage

Command Line Interface

The simplest way to use wave2midi is via the command line.

Basic Usage:

wave2midi /path/to/your/song.wav /path/to/output_directory/

With a custom configuration file:

wave2midi song.wav output/ --config my_config.json

With a different output BPM:

wave2midi song.wav output/ --bpm 140

Python API

You can also use the WaveToMIDIConverter class directly in your Python scripts.

from wave2midi import WaveToMIDIConverter

# Create a converter instance
converter = WaveToMIDIConverter()

# Convert a WAV file to MIDI stems
midi_files = converter.convert("input.wav", "output_dir/")

print(f"Created {len(midi_files)} MIDI files from stems.")

Configuration

You can configure the note detection and MIDI conversion parameters by creating a JSON file and passing it with the --config argument.

Here is an example config.json:

{
  "sample_rate": 22050,
  "pitch_detection_method": "pyin",
  "frame_length": 2048,
  "hop_length": 512,
  "fmin": 27.5,
  "fmax": 4186.01,
  "probability_threshold": 0.5,
  "min_note_duration": 0.1,
  "max_note_duration": 2.0,
  "velocity_scaling": 1.0,
  "output_bpm": 120
}

Limitations

• Complex polyphonic music may result in note detection inaccuracies.
• The quality of the source separation depends on the Demucs model and can vary.
• Percussive elements may not always translate perfectly to MIDI percussion tracks.
• Audio quality and recording conditions significantly affect conversion accuracy.

Acknowledgments

• The Demucs team for their excellent source separation model.
• The librosa development team for their audio analysis tools.
• The mido development team for their MIDI library.