HARMONY GENERATION:From Monophonic to Polyphonic Melodies

A neural harmony generator that generates four-part (SATB) harmonies from soprano melodies using a seq2seq LSTM architecture with attention mechanism (trained on JSB dataset - based on Bach's chorales).

Requirements

• Python dependencies: torch, numpy, matplotlib, mido
• GPU: optional (CPU training supported but slower)

Datset Download

• Dataset can be found at https://github.com/czhuang/JSB-Chorales-dataset
• Download and unzip Jsb16thSeparated.json
• Ensure the json file is in the same directory as the code file (harmony_generation_main.py)

Running the Model

Clone repository to your own machine

• Run main harmony generation model -> harmony_generation_main.py

Additional Info

• harmony_generation_main.py - main traning script
• HarmonyGeneration_Report.docx - report about the project
• Also contains training_curves.png, sample_temp.png (6 files - temperature 0.7 and temperature 1.0) and generated_sample_*.mid (3 files) generated from training

Repository Contents

• harmony_generation_main.py — main training script
• HarmonyGeneration_Report.docx — project report
• Also contains:
• training_curves.png — plot of training and validation performance
• sample_temp_0.7_*.png — generated output samples at temperature = 0.7 (3 files)
• sample_temp_1.0_*.png — generated output samples at temperature = 1.0 (3 files)
• generated_sample_*.mid — MIDI files generated from training (3 files)