Low-Complexity Acoustic Scene Classification Using Parallel Attention-Convolution Network

Yanxiong Li, Jiaxin Tan, Guoqing Chen, Jialong Li, Yongjie Si, Qianhua He

Abstract

This work presents an improved system submitted to Task 1 of the DCASE2023 challenge. We propose a method for low-complexity acoustic scene classification (ASC) using a Parallel Attention-Convolution Network (PACN). The architecture consists of four modules: pre-processing, local contextual information (LCI) extraction, global contextual information (GCI) extraction, and fusion. The proposed network is computationally efficient, capturing both fine-grained local details and long-range global relationships from audio clips. By integrating techniques such as knowledge distillation, data augmentation, and adaptive residual normalization, our method achieves a state-of-the-art accuracy of 56.10% on the DCASE2023 dataset. Remarkably, this is achieved with only 5.21k parameters and 1.44 million multiply-accumulate operations (MACs), outperforming the top systems in the challenge in both accuracy and complexity.

Introduction

Acoustic Scene Classification (ASC) involves identifying environments—such as an airport, a park, or a metro station—based on audio recordings. While deep learning has significantly improved ASC performance, deploying these models on resource-constrained edge devices (e.g., smartphones or hearing aids) remains difficult due to high computational demands.

Standard Convolutional Neural Networks (CNNs) are effective at capturing Local Contextual Information (LCI), such as short-term acoustic textures. However, they often struggle with Global Contextual Information (GCI), which represents the broader temporal structure of a scene. Previous attempts to combine these features often used serial processing, which increases latency and complexity. The PACN addresses this by processing local and global features in parallel.

Methodology

The PACN architecture is designed for extreme efficiency, utilizing four distinct modules to process audio data:

  1. Pre-processing: Raw audio is converted into log Mel-spectrograms. Blueprint Separable Convolutions (BSConv) and max pooling are used to reduce the initial feature size.
  2. LCI Extraction: This branch utilizes BSConv and Global Response Normalization (GRN) to focus on fine-grained, channel-wise local details.
  3. GCI Extraction: This branch employs a Transformer-style architecture with Multi-Head Attention to capture long-range relationships between different time frames in the audio.
  4. Fusion: The outputs from the parallel branches are concatenated. A "channel shuffle" operation is used to mix information before a final fully connected (FC) layer produces the classification.
The PACN architecture consists of four main modules: Pre-processing, LCI extraction, GCI extraction, and Fusion.

To maximize performance within a tiny parameter budget, the authors employed Knowledge Distillation (KD), where a larger "teacher" model guides the PACN "student." They also used Adaptive Residual Normalization (ARN) to ensure the model remains robust across different recording devices and hardware.

Experimental Setup

The model was evaluated on the TAU Urban Acoustic Scene 2022 Mobile development dataset (TAU22). This dataset includes over 230,000 audio clips across 10 scene categories, recorded in 10 different cities using 9 different devices (3 real, 6 simulated). This diversity tests the model's ability to generalize to unseen hardware and environments.

Results and Discussion

The PACN sets a new benchmark for efficiency in the DCASE2023 Task 1 challenge. It achieved an accuracy of 56.10%, surpassing top-performing systems like CP-Mobile (54.66%) and TF-SpeNet (53.90%).

Rank matrix and Critical Difference (CD) diagram comparing the proposed PACN (Ours) against DCASE baselines and other state-of-the-art models.

As shown in the rank matrix and Critical Difference (CD) diagram, the proposed method consistently ranks first and is statistically superior to the baselines. Despite its high performance, the model uses only 5,210 parameters and 1.44 million MACs, making it exceptionally suitable for real-time deployment on low-power intelligent audio terminals.

Conclusion

The Parallel Attention-Convolution Network demonstrates that ultra-low complexity models can achieve state-of-the-art results in acoustic scene classification. By leveraging parallel branches for local and global context and utilizing advanced training strategies like knowledge distillation, the PACN provides a robust solution for edge-based audio AI. Future work will explore mutual-attention mechanisms to further enhance feature integration between the parallel branches.

Code is available at: https://github.com/Jessytan/Low-complexity-ASC

Summary updated

BibTeX
@misc{li2024low,
  author = {Yanxiong Li and Jiaxin Tan and Guoqing Chen and Jialong Li and Yongjie Si and Qianhua He},
  title = {{Low-Complexity Acoustic Scene Classification Using Parallel Attention-Convolution Network}},
  year = {2024},
  eprint = {2406.08119},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.08119},
}
Figure