This is exactly the MIREX-2013 submission for the Audio Classification (Test/Train) 
task of "Audio Latin Music Genre Classification". This submission won the 1st place. 

The code implements the algorithm:

Pikrakis, A., “A deep learning approach to rhythm modeling with applications”, 
6th International Workshop on Machine Learning  and Music (ÌML 2013,), 
held in conjunction with the European Conference on Machine Learning and 
Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD 2013), 
Prague, Czech Republic, September 23, 2013 (to appear).

The perfomance of this algorithm, including its limitations, was further studied at:
B. L. Sturm, C. Kereliuk and A. Pikrakis, "A closer look at deep learning neural networks with low-level spectral 
periodicity features," 2014 4th International Workshop on Cognitive Information Processing (CIP), Copenhagen, 2014, 
pp. 1-6.


The submission consists of MATLAB code. It does not assume that any specific 
MATLAB libraries are present in the system where the code will run.

--------------------------------------------
INSTALLATION INSTRUCTIONS
--------------------------------------------
You only need to download and unzip to a single folder.

------------------
HOW TO RUN
------------------
A) Feature extraction:
Use the m-file extractFeatures.m
Example:
extractFeatures('.\tmp','C:\Users\aggelos\Documents\MATLAB\mirEX2013\tmp\featureExtractionListFile.txt');
In the end of the feature extraction stage, the scratch folder ('.\tmp') will contain one feature file per audio file. 
Arguments of extractFeatures are:
* 1st argument: PathToScratchFolder (string) is the full path to the scratch folder. 
* 2nd argument: PathToFeatureExtractionListFile is the full path to the txt file that contains one filename per row.
* IT IS IMPORTANT THAT THE FILENAMES OF THE AUDIO FILES ARE UNIQUE.

B) Training:
Use the m-file TrainingAlgorithm.m.
Example:
TrainingAlgorithm('.\tmp','.\tmp\TrainingListFile.txt');
Arguments of TrainingAlgorithm are:
* PathToScratchFolder (string) is the full path to the corresponding folder. 
* TrainingSetListFile (string) is the full path to the corresponding file. 
* The TrainingAlgorithm.m file reads feature files from the scratch folder and 
writes its output to the TrainingData.mat file in the scratch folder.

C) Classification:
Use the m-file ClassifyAlgorithm.m
Example:
ClassifyAlgorithm('.\tmp','.\tmp\TestListFile.txt','.\tmp\OutputFile.txt');
Arguments of ClassifyAlgorithm are:
* PathToScratchFolder (string) is the full path to the corresponding folder. 
* TestingSet (string) is the full path to the corresponding file. 
* The ClassifyAlgorithm.m file writes its output to the OutputListFile.
* outputListFile (string) is the full path to the corresponding output file. 

All files conform to the format of the guidelines that have been set by the task organizers.

--------------------------------------
DISKPACE REQUIREMENTS
--------------------------------------
The training algorithm stores in the scratch folder one feature file per audio recording. For the MIREX-2013 task, the size of each feature file was approximately 150 KB. Assuming 3500 recordings, the storage requirements are around 500 MB. The rest of the files that the training algorithm creates are negligible with respect to size.

-------------------------------------
RUNTIME REQUIREMENTS
-------------------------------------
The most computationally demanding stage is the feature extraction stage. Assuming 3500 recordings, 3 min each, it is expected that it will take around 8 hours for the feature extraction stage to complete.

The training stage will consume around 5 hours of processing time on a standard laptop.

The classification stage is trivial. It should take a few minutes to complete.

-------------------
MEMORY REQUIREMENTS
-------------------
Due to the use of a deep learning architecture, the algorithm is memory demanding. However, we do not expect any memory problems to emerge. If this is the case, a memory efficient but radically slower version can be made available on demand.