Note:: This repo is a work in progress. This README file is the ongoing documentation, and all scripts are not yet uploaded/up-to-date.
This repository contains the accompanying code for the paper:
Helena Cuesta & Emilia Gómez. (2022). Voice Assignment in Vocal Quartets Using Deep Learning Models Based on Pitch Salience. Transactions of the International Society for Music Information Retrieval (TISMIR), 5(1), 99–112. DOI: http://doi.org/10.5334/tismir.121.
This is the accompanying code repository for the paper mentioned above. It currently contains the trained models for
Voice Assignment (VA), namely voas_cnn and voas_clstm and the associated code to use them with pre-extracted
pitch salience representations -- see Priors for information on how to obtain such representations.
Additionals scripts to reproduce/extend the examples from the Synth-salience Choral Set (SSCS), the synthetic dataset considered in this work, will be provided soon.
The dataset can be downloaded following
this link, and the
training/validation/test data splits are available in data/data_splits_hpc.json.
These VA models take polyphonic pitch salience representations as input, and they assume four different sources. In our paper, we specifically propose a framework that combines the output of Late/Deep (https://github.com/helenacuesta/multif0-estimation-polyvocals, model3) with the proposed VA models---VoasCNN and VoasCLSTM, on vocal quartets.
To use the trained models, please run the predict_on_salience.py script with the following parameters:
--model: Model to use for prediction: voas_clstm | voas_cnn
--saliencefile: Path to the input salience file. It expects a npy files.
--saliencefolder: Path to the folder with salience files.
--outputpath: Path to the folder to store the results. If nothing is provided, results will be stored in the same folder of the input(s).")