README

2016.06.24
This is sourcecode of DCASE Challenge 2016 Task 4. 
DCASE Challenge 2016 Task 4. http://www.cs.tut.fi/sgn/arg/dcase2016/task-audio-tagging
Using Mel + DNN achieve 19% EER using 4 fold cross validation on development dataset. 

THIS CODE IS BSD LISCENCED. 
CONTACT: QIUQIANG KONG, q.kong@surrey.ac.uk
CVSSP, UNIVERSITY OF SURREY

======================================

Pre-requiste:
pip install librosa	extract features
pip install hat 	deep learning toolbox, https://qiuqiangkong.github.io/Hat

Usage:
------ Preparation ------
* modify the paths of dataset in config.py

------ Development Cross Validation ------
* run prepare_dev_data.py 		Calculate features for development data
* run main_dev_dnn.py 			Train model on cross validation of development data
* run main_dev_recognize.py 		Generate results and evaluate on cross validation data

------ Generate Results of Private Data ------
* run prepare_eva_data.py	Calculate features for evaluation data
* run main_dnn_eva.py		Train model on all development data
* run main_eva_recognize.py	Generate Results of private data