This repo is provided as a supplement to the paper Lyrics and Vocal Melody Generation conditioned on Accompaniment. It has been currently accepted at the 2nd Workshop on NLP for Music and Spoken Audio NLP4MuSA 2021.
It includes:
- All code to recreate the results of the paper
- The full instrumental dataset
dataset.parquetand the reduced chord datasetdataset_chords.parquet, as well as their token vocabularies - A few generated samples and their corresponding lyrics
This work is the continuation of my Master's thesis (MEng in ECE at NTUA)
The provided code includes scripts to:
- pre-process MIDI tracks and bring them to the used text event format
- build the full and chord reduced dataset (as well as the token vocabularies) from the LMD matched dataset
- train the vanilla encoder-decoder architecture
- train the decoupled architecture
- warm-start and finetune the lyrics language model
- adjust Mellotron in order to use our text event format
- train all models and generate samples on Colab (Jupyter Notebooks)
Since the git lfs free data limit has been reached, here is a link for a Google Drive folder containing the necessary data.