Acoustic Scene Classification by Implicitly Identifying Distinct Sound Events

Introduction

This is the code for the paper arxiv:1904.05204 (download)

The paper is submitted to Interspeech 2019.

install

• create env and activate

conda create -n distinct_events_asc python=3.6 pip
source activate distinct_events_asc

• install pytorch torchvision from pytorch channel

conda install pytorch torchvision -c pytorch

• install requirements with pip

pip install -r requirements.txt

• if want to run jupyter notebook examples, install a kernelspec for env

conda install jupyter ipykernel
python -m ipykernel install --user --name distinct_events_asc --display-name 'python3.6(distinct_events_asc)'

data_manager

NOTE: before use, config data_manager.cfg properly

• create file data_manager.cfg under data_manager/
• specify dev_path to point to dcase2018 Task1 SubTaskB dataset

[DEFAULT]

[dcase18_taskb]
dev_path = /PathTo.../dcase2018_baseline/task1/datasets/TUT-urban-acoustic-scenes-2018-mobile-development

[logmel]
sr = 44100
n_fft = 1764
hop_length = 882
n_mels = 40
fmax = 22050

• extract and store feature in .h5 file

# generate .h5 files under data_manager/data_h5 
python data_manager/dcase18_taskb.py
# generate scaler .h5 under data_manager/data_h5
python data_manager/taskb_standrizer.py

Experiments

• open jupyter notebook or jupyter lab
• run experiments notebooks under jupyter_exp/

Citation

• BibLatex

@online{1904.05204,
Author = {Hongwei Song and Jiqing Han and Shiwen Deng and Zhihao Du},
Title = {Acoustic Scene Classification by Implicitly Identifying Distinct Sound Events},
Year = {2019},
Eprint = {1904.05204},
Eprinttype = {arXiv},
}

• Bibtex

@misc{1904.05204,
Author = {Hongwei Song and Jiqing Han and Shiwen Deng and Zhihao Du},
Title = {Acoustic Scene Classification by Implicitly Identifying Distinct Sound Events},
Year = {2019},
Eprint = {arXiv:1904.05204},
}