Skip to content
forked from yzspku/CQTNet

LEARNING A REPRESENTATION FOR COVER SONG IDENTIFICATION USING CONVOLUTIONAL NEURAL NETWORK. ICASSP2020

License

Notifications You must be signed in to change notification settings

silvadirceu/CQTNet

 
 

Repository files navigation

CQTNet

LEARNING A REPRESENTATION FOR COVER SONG IDENTIFICATION USING CONVOLUTIONAL NEURAL NETWORK. ICASSP2020

Environment

python -- 3 pytorch -- 1.0 librosa -- 0.63

Dataset

Second Hand Songs 100K (SHS100K), which is collected from Second Hand Songs website.

Generate CQT

You can utilize "gencqt.py" to get CQT features from your own audio.

Train

python main.py multi_train --model='CQTNet' --batch_size=32 --load_latest=False --notes='experiment0'

Test

python main.py test --model='CQTNet' --load_model_path = 'check_points/CQTNet.pth'

Paramters

https://drive.google.com/file/d/1Rv-NuiAKW2rUlNZj8SOs2Iqidqkx30M8/view?usp=sharing

Spectrum Augmentation

After using Spectrum Augmentation in training stage, the model performance has a great improvement.

Specaugment: A simple data augmentation method for automatic speech recognition.

Dataset MAP
YouTube350 0.933
Covers80 0.860
Mazurkas 0.933
SHS100K-TEST 0.71

About

LEARNING A REPRESENTATION FOR COVER SONG IDENTIFICATION USING CONVOLUTIONAL NEURAL NETWORK. ICASSP2020

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 98.6%
  • Python 1.4%