machine-listeners-valencia / dcase2021-task1

Codes related to DCASE2021 Task 1 - Acoustic Scene Classification

DCASE2021-Task1

Codes related to DCASE2021 Task 1 - Acoustic Scene Classification

Acoustic Scene Classification is a machine listening task whose goal is to map an audio clip into a pre-defined scene target.

The mismatch devices consideration appears when the audio clips have been recorded with different devices.

DCASE2020 submission

The submission was carried out using the squeeze-excitation network presented in https://ieeexplore.ieee.org/abstract/document/9118879

For more details please of the submission please we refer to the technical report: http://dcase.community/documents/challenge2020/technical_reports/DCASE2020_Naranjo-Alcazar_34_t1.pdf

The best system obtained 65.1% accuracy being 54.1% the baseline in development stage

In evaluation stage, the system performed 61.9% accuracy being 51.4% the baseline.

If use this papers, please cite them:

@article{naranjo2020acoustic,
  title={Acoustic scene classification with squeeze-excitation residual networks},
  author={Naranjo-Alcazar, Javier and Perez-Castanos, Sergi and Zuccarello, Pedro and Cobos, Maximo},
  journal={IEEE Access},
  volume={8},
  pages={112287--112296},
  year={2020},
  publisher={IEEE}
}

@techreport{naranjo2020task,
  title={TASK 1 DCASE 2020: ASC WITH MISMATCH DEVICES AND REDUCED SIZE MODEL USING RESIDUAL SQUEEZE-EXCITATION CNNS},
  author={Naranjo-Alcazar, Javier and Perez-Castanos, Sergi and Zuccarello, Pedro and Cobos, Maximo},
  year={2020},
  institution={DCASE2020 Challenge, Tech. Rep}
}

First modifications

Due to submissions presented in the DCASE2020 edition some modifications to our model. Training proceudre (callbacks, epochs and mixup remain the same)

Modification	Frequency bins	Accuracy (%)
Same model, focal loss	64	65.56
Removing Dense 100 units layer, flatten reshape, focal loss	64	63.57
Removing Dense 100 units layer, global average reshape, focal loss	64	66.36
Removing Dense 100 units layer, global average reshape, no freq pooling, focal loss	64	67.81
Removing Dense 100 units layer, global average reshape, no freq pooling, focal loss	128	67.00

Discussion

• focal loss improves system's performance
• no reducing the frequency bins also improves system's performance
• adding more frequency bins does not improve system's performance if doing the same procedure
• Flatten and Dense layers are more prone to overfitting in this scenario

Run Code

• This repo assumes that it has been saved as a subdirectory of HOME

• You only have to configure the config.py file as desired to launch trainings