Timbre Analysis of Music Audio Signals with Convolutional Neural Networks: Musical instrument recognition

This repository contains code for musical instrument recognition experiments for the paper entitled "Timbre Analysis of Music Audio Signals with Convolutional Neural Networks".

We provide the code for data preprocessing, training and evaluation of our approach.

Dataset

We use IRMAS dataset. Please, download the dataset from [MTG website] (www.mtg.upf.edu/download/datasets/irmas) and change the paths to the training and testing splits at the settings file: ./experiments/settings.txt

Preprocessing

Before to run the preprocessing script, please, create specific folder for every model which you would like to preprocess the data for. The following notation is using (consult ./experiments/settings.txt): $IRMAS_HOME_DIRECTORY/IRMAS-TrainingData-Features/model_name and $IRMAS_HOME_DIRECTORY/IRMAS-TestingData-Features/model_name, where model_name is referenced to one of the model filenames in ./experiments/models folder. Usage:

python preprocessing.py -m model_name

Currently supported models are han16, singlelayer, multilayer which are referenced in the paper as han16, single-layer and multi-layer models respectively.

Training

Usage:

python training.py -m model_name -o optimizer_name [-l]

The option -l states for loading data into RAM at the beginning of the experiment instead of reading it batch-by-batch from the disk.

Evaluation

Usage:

python evaluation.py -m model_name -w /path/to/weights/file.hdf5 -s evaluation_strategy

The weights for the reported models can be found at ./weights/model_name folder.

Evaluation strategies are: The s1 strategy computes a mean activation through whole audio excerpt and apply identification threshold The s2 strategy computes sum of activations, normalize it by dividing by maximum activation.

Reference

References are coming.