[Paper Review] Integrating the Data Augmentation Scheme with Various Classifiers for Acoustic Scene Modeling

Hangting Chen, Zuozhen Liu+3 more|arXiv (Cornell University)|Jul 15, 2019

Music and Audio ProcessingComputer Science12 references67 citations

TL;DR

The paper presents a data augmentation framework using GAN-based methods (ACGAN and CVAE/ACGAN) integrated with multiple ASC classifiers (FBank-FCNN, Scalogram-DCNN, and hybrids) and ensemble voting to achieve over 85% accuracy on DCASE2019 Task1A fold-1.

ABSTRACT

This technical report describes the IOA team's submission for TASK1A of DCASE2019 challenge. Our acoustic scene classification (ASC) system adopts a data augmentation scheme employing generative adversary networks. Two major classifiers, 1D deep convolutional neural network integrated with scalogram features and 2D fully convolutional neural network integrated with Mel filter bank features, are deployed in the scheme. Other approaches, such as adversary city adaptation, temporal module based on discrete cosine transform and hybrid architectures, have been developed for further fusion. The results of our experiments indicates that the final fusion systems A-D could achieve an accuracy higher than 85% on the officially provided fold 1 evaluation dataset.

Motivation & Objective

• Motivate robust acoustic scene classification (ASC) under domain shifts (e.g., unseen cities) using data augmentation.
• Develop and compare ASC classifiers using different feature representations (Mel filter bank, scalogram) and architectures (1D DCNN, 2D FCNN, Inception-based hybrids).
• Demonstrate fusion/ensembling strategies to improve final ASC performance.
• Assess the impact of auxiliary GAN-based augmentation (ACGAN, CVAE/ACGAN) and adversarial domain adaptation on ASC performance.

Proposed method

• Employ ACGAN and CVAE/ACGAN data augmentation to generate synthetic acoustic feature maps with scene labels.
• Use two primary classifiers: (i) FBank-FCNN (10-class) and (ii) Scalogram-DCNN (10-class); both trained with and without augmentation.
• Incorporate additional classifiers: DCT-based temporal module, Adversarial City Adaptation, and hybrid Inception-based networks (IncepLSTM/IncepGRU) to enable model fusion.
• Explore ensemble strategies (average and weighted voting) to combine multiple classifiers for final predictions.
• Train and evaluate on fold-1 DCASE2019 setup; perform early stopping and hyperparameter tuning using a validation split; retrain best systems on full development data for submission.

Experimental results

Research questions

• RQ1 How does GAN-based augmentation (ACGAN, CVAE/ACGAN) affect ASC performance across different feature representations (FBank vs scalogram) and architectures?
• RQ2 What is the impact of adversarial city adaptation and DCT-based temporal modules on generalization to unseen domains/cities?
• RQ3 Can ensemble voting (average vs weighted) of diverse ASC models surpass single-model performance, and by how much?
• RQ4 Which combinations of features, augmentation schemes, and classifier architectures yield the highest fold-1 accuracy on DCASE2019 Task1A?

Key findings

• Data augmentation with ACGAN or CVAE/ACGAN improves ASC accuracy across feature types and classifiers (gains of ~0.5–4% observed).
• Scalogram-based features with ave-diff channels generally outperform left-right channels by ~3–5%.
• Best single systems reach up to 84.28% with scalogram-aveDiff and CVAE/ACGAN-DCNN; CVAE/ACGAN can outperform ACGAN in some setups.
• Hybrid and adversarial approaches provide complementary gains, but combining all enhancements does not always improve performance; DCT can help in some configurations.
• Fusion systems achieve final accuracies of 85.07%–85.28% on fold-1 evaluation, exceeding 85% with proper weighting and averaging.