Projects

Re-MOVE
[Github]
Re-MOVE (Reduced MOVE) is a neural network model that is trained with embedding distillation techniques to further improve both the accuracy and the scalability aspects of MOVE. This repository contains the training/evaluation code and the pre-trained model weights.
Keywords: Python, PyTorch, embedding distillation, knowledge distillation, neural network pruning, dimensionality reduction, metric learning.

MOVE
[Github]
MOVE, or musically-motivated version embeddings, is a convolutional neural network model that demonstrates state-of-the-art performance in version identification task. It is designed to achieve invariances against changes in pitch transpositions, tempo/timing, and structure.
Keywords: Python, PyTorch, invariant representations, neural network encoder, attention, metric learning, data augmentations.

Da-TACOS
[Github]
Da-TACOS is a dataset for cover song identification and understanding, which contains various pre-extracted features for 25k songs. The feature extraction and benchmarking are performed using the “acoss” framework that is specifically designed for this task.
Keywords: Python, audio signal processing, open science.

PhonationRT
[Github]
PhonationRT is a visual feedback prototype for learning/teaching phonation modes in singing.
Keywords: C++, Qt Creator, Weka.