A deep learning project for musical interval recognition using PyTorch. This system generates synthetic musical intervals, processes them into mel spectrograms, and trains a CNN model to recognize different musical intervals.

intervals-ai/
│
├── intervals_ai/
│   ├── dataset/
│   │   ├── dataset.py         # Dataset handling
│   │   └── preprocessor.py    # Audio preprocessing
│   ├── generation/
│   │   ├── config.py          # Configuration classes
│   │   ├── intervals.py       # Interval generation
│   │   └── notes.py          # Note generation utilities
│   ├── models/
│   │   ├── cnn.py            # CNN model architecture
│   │   ├── evaluation.py     # Model evaluation
│   │   └── trainer.py        # Training logic
│   ├── utils/
│   │   ├── audio.py          # Audio utilities
│   │   └── visualization.py  # Visualization utilities
│   └── main.py               # Main training script
│
├── generated_data/           # Generated audio samples
│   └── interval_*.wav        # Generated interval audio files
│
├── experiments/             # Training runs
│   └── run_YYYYMMDD_HHMMSS/ # Timestamped experiment directory
│       ├── checkpoints/     # Model checkpoints
│       ├── plots/           # Generated visualizations
│       └── stats/           # Training statistics and metrics
│
├── README.md
├── pyproject.toml          # Poetry dependencies
└── poetry.lock            # Poetry lock file

1. Install Python 3.10 or newer
2. Install Poetry (dependency management):

   curl -sSL https://install.python-poetry.org | python3 -

3. Clone the repository:

   git clone https://github.com/mj300405/intervals_ai.git
   cd intervals-ai

4. Install dependencies:

   poetry install

1. Activate the virtual environment:

   poetry shell

2. Run the training script:

   python intervals_ai/main.py

The script will:

1. Generate a dataset of 5000 interval samples (if not already present)
2. Train a CNN model for 100 epochs
3. Save results in a timestamped experiment directory

Each training run creates a timestamped directory under experiments/:

experiments/run_YYYYMMDD_HHMMSS/
├── checkpoints/
│   ├── best_model.pt           # Best performing model
│   └── checkpoint_epoch_*.pt   # Regular checkpoints
│
├── plots/
│   ├── confusion_matrix.png    # Model performance matrix
│   ├── training_history.png    # Loss and accuracy curves
│   └── audio_features_*.png    # Audio visualizations
│
└── stats/
    ├── run_config.json         # Run configuration
    ├── training_history.json   # Detailed training metrics
    ├── evaluation_results.json # Test set results
    ├── metrics_summary.txt     # Performance summary
    └── training_summary.txt    # Training overview

• Input: Mel spectrograms (128 mel bands)
• Architecture: CNN with 3 convolutional layers
• Training:
  • 5000 total samples
  • 70/15/15 train/val/test split
  • 150 epochs maximum
  • Early stopping with 15 epochs patience
  • Learning rate: 0.001
  • Batch size: 32

The system generates intervals using these parameters:

• Sample rate: 44100 Hz
• Duration: 2.0 seconds
• Note gap: 0.2 seconds
• Frequency range: 55 Hz (A1) to 1760 Hz (A6)
• Supported intervals:
  • Minor/Major Second
  • Minor/Major Third
  • Perfect Fourth
  • Tritone
  • Perfect Fifth
  • Minor/Major Sixth
  • Minor/Major Seventh
  • Octave

1. Training Plots:

   • Loss curves (training and validation)
   • Accuracy curves (training and validation)
   • Confusion matrix

2. Audio Visualizations:

   • Waveform
   • Spectrogram
   • Mel spectrogram

Main dependencies include:

• PyTorch (deep learning)
• librosa (audio processing)
• numpy (numerical computations)
• matplotlib (visualization)
• soundfile (audio I/O)
• scikit-learn (evaluation metrics)

For a complete list, see pyproject.toml.

MIT License

• First run will generate the dataset, which may take some time
• Subsequent runs will reuse the existing dataset
• Each run creates a new experiment directory
• GPU will be used if available, otherwise CPU
• Training progress is saved regularly, allowing for recovery