Jazz Music Generation

Overview

This project demonstrates how to generate jazz music using a Long Short-Term Memory (LSTM) network. The model is trained on professional jazz performances and learns to improvise solos based on input sequences.

Features

• Uses an LSTM-based recurrent neural network to generate jazz solos.
• Trained on preprocessed musical data in the form of numerical sequences.
• Outputs MIDI sequences that can be converted into playable jazz solos.
• Implements deep learning techniques with TensorFlow and Keras.

Dataset

The dataset consists of musical values extracted from jazz performances:

• X: A (m, 30, 90) shaped array containing input sequences, where:
  • m is the number of training examples.
  • T_x = 30 represents the sequence length.
  • 90 is the number of unique possible values (one-hot encoded notes/chords).
• Y: The same as X, but shifted one step to the left for supervised training.
• n_values: The number of unique values in the dataset (90 total).
• indices_values: A mapping of integer indices to musical values.

Dependencies

To run this project, install the following Python packages:

pip install tensorflow numpy music21 matplotlib

How to Run

1. Clone this repository and navigate to the project directory.
2. Open and execute the Jupyter Notebook Improvise_a_Jazz_Solo_with_an_LSTM_Network_v4.ipynb.
3. Train the model on the dataset.
4. Generate new jazz solos using the trained model.

Model Architecture

• LSTM Layers: Capture sequential dependencies in musical notes.
• Dense Layers: Convert LSTM outputs into probabilities over possible notes.
• Softmax Activation: Determines the next note probabilistically.
• Categorical Crossentropy Loss: Optimizes the model during training.

Results

The trained model can generate realistic jazz improvisations by predicting note sequences in real-time. The generated MIDI files can be converted into playable formats for evaluation.

Future Improvements

• Expand dataset with more jazz recordings.
• Fine-tune hyperparameters for improved realism.
• Implement real-time MIDI generation and playback.

Credits

This project is inspired by deep learning applications in music generation. The implementation is based on research in sequence modeling with LSTMs.

License

This project is open-source and available for use under the MIT License.

---

Author: Chirag N Vijay