Entry to the DCASE 2018 task 4 (Large-scale weakly labeled semi-supervised sound event detection in domestic environments) challenge

Based on Convolutional Recurent Neural Network. Tensorflow / Keras

About the challenge

To get information about the challenge, please visit its Website. The purpose was to propose a solution to a classification and localization task of multiple sound events in domestics environments. 10 classes describe these events and an audio file can be tagged with multiple class, sometimes overlapping. (weak annotation)

Requirements

conda create -n dcase python=3.5 pip
conda activate dcase
pip install scikit-learn
pip install dcase_util
pip install sed_eval
pip install tensorflow-gpu
pip install keras

Usage

Main arguments

• --normalizer Normalization method to apply on data before training [file_MinMax | global_MinMax | file_Mean | global Mean | file_standard | global_standard | unit ]
• --output_model Basename of the models to save
• --meta_root Path to meta files location (root)
• --features_root Path to feature files location (root)
• -uid Use the unlabel in domain subset to retrain first model and perform adaptation
• -retrain Force the program to retrain the model even if it already exist
• -w If set, display hidden warnings

conda activate dcase
python main.py --output_model results/model.name --meta_root path/to/meta/root --features_root path/to/features/root -uid

Weighted Gate Recurent Unit (WGRU)

To achieve the current score, a Weighted Gate Recurrent Unit (WGRU) was used in parallel with a classic Gate Recurrent Unit (GRU). The GRU has a good performance when it comes to "stationary" sound localization whereas the WGRU perform better with punctual sounds such as speech, dog, alarm bell and cat.

Weight injection in the GRU layer	Weight impact on localization