Improvise a Jazz Solo with an LSTM Network 🎷🎶

Project Overview

This project focuses on generating jazz music using deep learning, specifically an LSTM-based (Long Short-Term Memory) neural network. The goal is to create a system capable of improvising jazz solos in a style similar to a trained dataset.

Objectives:

• Apply an LSTM to a music generation task.
• Generate original jazz music based on learned patterns.
• Utilize the Functional API of TensorFlow/Keras to design and train a complex sequential model.

Dataset

The model is trained on a corpus of jazz music preprocessed into sequences of musical "values." These values represent musical notes or chords encoded as one-hot vectors. The dataset is formatted as:

• X: A (m, T_x, 90) array where each sequence contains 30 notes (T_x = 30) with 90 possible values.
• Y: The labels shifted by one time step to predict the next value.

Methodology

1. Preprocessing:

   • Musical pieces are converted into sequences of values.
   • Data is split into inputs (X) and corresponding outputs (Y).

2. Model Architecture:

   • Uses an LSTM network with a hidden state size of 64.
   • Inputs are reshaped to match the LSTM requirements.
   • A dense layer with softmax activation is used for output predictions.

3. Training:
   • Trains on snippets of 30 musical values.
   • Optimizes to predict the next note in a sequence using categorical cross-entropy loss.

4. Generation:
   • Implements a custom sequence generation loop.
   • At each step, the predicted note is fed as input to generate the next note iteratively.

Tools and Libraries

• TensorFlow/Keras
• music21 for handling MIDI data.
• Supporting modules like preprocess.py and music_utils.py for data preparation.

Results

After training, the model generates jazz solos by predicting notes iteratively. The output can be converted back into MIDI format for playback.

Usage

Prerequisites:

• Python 3.x
• Required libraries: TensorFlow, NumPy, music21, Matplotlib

Steps to Run:

1. Clone the repository or download the project files.
2. Install the dependencies:

   pip install -r requirements.txt

3. Run the notebook:

   jupyter notebook Improvise_a_Jazz_Solo_with_an_LSTM_Network.ipynb

4. Train the model on the provided dataset.
5. Use the trained model to generate jazz solos.

Outputs:

• Generated MIDI files can be played to listen to the jazz solo.

Future Work

• Extend the model to handle other genres of music.
• Experiment with more advanced architectures like transformers for better results.
• Incorporate real-time generation and visualization of music.