Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Acoustic Scene Classification in Low-Resource Conditions

Zhi Chen, Yun-Fei Shao, Yong Ma, Mingsheng Wei, Le Zhang, Wei-Qiang Zhang

TL;DR

This work addresses acoustic scene classification (ASC) under low-resource conditions and cross-device variability. It proposes DS-FlexiNet, a lightweight architecture that fuses depthwise separable convolutions with ResNet-like residual connections, augmented by Quantization Aware Training (QAT), Auto Device Impulse Response (ADIR), and Freq-MixStyle (FMS), plus Knowledge Distillation (KD) from twelve teacher models. Key innovations include Residual Normalization with RN(x)=\lambda\cdot x+IN(x) to capture device-specific features, frequency-based augmentation via FMS, energy-aware DIR simulation via ADIR, and a soft-teacher fusion scheme that combines logits across teachers; experimental results on the TAU Urban Acoustic Scene 2022 Mobile Development dataset demonstrate improved robustness and efficiency. The approach enables robust ASC deployment on mobile and embedded devices, addressing inter-device variations and resource constraints.

Abstract

Acoustic Scene Classification (ASC) identifies an environment based on an audio signal. This paper explores ASC in low-resource conditions and proposes a novel model, DS-FlexiNet, which combines depthwise separable convolutions from MobileNetV2 with ResNet-inspired residual connections for a balance of efficiency and accuracy. To address hardware limitations and device heterogeneity, DS-FlexiNet employs Quantization Aware Training (QAT) for model compression and data augmentation methods like Auto Device Impulse Response (ADIR) and Freq-MixStyle (FMS) to improve cross-device generalization. Knowledge Distillation (KD) from twelve teacher models further enhances performance on unseen devices. The architecture includes a custom Residual Normalization layer to handle domain differences across devices, and depthwise separable convolutions reduce computational overhead without sacrificing feature representation. Experimental results show that DS-FlexiNet excels in both adaptability and performance under resource-constrained conditions.

Improving Acoustic Scene Classification in Low-Resource Conditions

TL;DR

This work addresses acoustic scene classification (ASC) under low-resource conditions and cross-device variability. It proposes DS-FlexiNet, a lightweight architecture that fuses depthwise separable convolutions with ResNet-like residual connections, augmented by Quantization Aware Training (QAT), Auto Device Impulse Response (ADIR), and Freq-MixStyle (FMS), plus Knowledge Distillation (KD) from twelve teacher models. Key innovations include Residual Normalization with RN(x)=\lambda\cdot x+IN(x) to capture device-specific features, frequency-based augmentation via FMS, energy-aware DIR simulation via ADIR, and a soft-teacher fusion scheme that combines logits across teachers; experimental results on the TAU Urban Acoustic Scene 2022 Mobile Development dataset demonstrate improved robustness and efficiency. The approach enables robust ASC deployment on mobile and embedded devices, addressing inter-device variations and resource constraints.

Abstract

Acoustic Scene Classification (ASC) identifies an environment based on an audio signal. This paper explores ASC in low-resource conditions and proposes a novel model, DS-FlexiNet, which combines depthwise separable convolutions from MobileNetV2 with ResNet-inspired residual connections for a balance of efficiency and accuracy. To address hardware limitations and device heterogeneity, DS-FlexiNet employs Quantization Aware Training (QAT) for model compression and data augmentation methods like Auto Device Impulse Response (ADIR) and Freq-MixStyle (FMS) to improve cross-device generalization. Knowledge Distillation (KD) from twelve teacher models further enhances performance on unseen devices. The architecture includes a custom Residual Normalization layer to handle domain differences across devices, and depthwise separable convolutions reduce computational overhead without sacrificing feature representation. Experimental results show that DS-FlexiNet excels in both adaptability and performance under resource-constrained conditions.
Paper Structure (19 sections, 5 equations, 3 figures, 5 tables)

This paper contains 19 sections, 5 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Network Architecture
  4. Residual Normalization
  5. Data Augmentation
  6. Freq-MixStyle
  7. Auto Device Impulse Response
  8. Teacher Models Fusion
  9. Experiment setup
  10. Dataset
  11. Experimental Preparation
  12. Quantization Aware Training
  13. Results and Analysis
  14. Proposed Models
  15. Performance Optimization and Resource Utilization
  16. ...and 4 more sections

Figures (3)

  • Figure 1: Network Architecture
  • Figure 2: Energy Distribution Analysis of Audio in the TAU22 Dataset.
  • Figure 3: The confusion matrix classified by different dimensions: MS:metro_station, PS: public_square, SM: shopping_mall, SP: street_pedestrian, ST: street_traffic. The comparison experiment is based on the sm4 model.