The phoneme classification code for EUSIPCO 2017 paper review:

Timbre Analysis of Music Audio Signals with Convolutional Neural Networks

1. Clone this repository
2. Download Jingju a capella singing dataset from http://doi.org/10.5281/zenodo.344932
3. Change dataset_path variable in parameters.py to locate the above dataset
4. Install dependencies (see below)
5. Choose dataset in parameters.py to run experiment on dan or laosheng dataset
6. Run experiment by `python doPhonemeClassification.py'

1. Execute the steps 1, 2, 3 in Steps for reproducting the experiment results
2. Choose dataset and am variables in parameters. Example, dataset='qmLonUpfLaosheng' and am='cnn' means we would like to extract the laosheng features for convolutional neural networks (proposed, Choi models).
3. Run python phonemeSampleCollection.py to extract the mel bands features
4. Code for extracting features for MLP model is not included.

1. Download pre-computed mel-bands features from http://doi.org/10.5281/zenodo.344935
2. Create a folder named trainingData in the root of this repository, then put all '.pickle.gz` feature files into this folder
3. If you don't want to download the pre-computed features, please follow Steps for calculating the mel bands features
4. The model training code are located in pretrainedDLModels folder. keras_cnn* code is for training CNN models (proposed and Choi modes). keras_dnn* code is for training MLP model
5. To train GMM models, please set am='gmm' in parameters.py, then execute steps 1, 2 in Steps for calculating the mel bands features

Steps for reproducting the experiment results requires below packages:

python2 numpy scipy scikit-learn matplotlib essentia

Steps for calculating the mel bands features requires below packages:

python2 numpy scipy scikit-learn essentia

Steps for training proposed, Choi, MLP and GMM models requires below packages:

python2 numpy scipy scikit-learn essentia keras theano hyperot

Affero GNU General Public License version 3