Research map / Architectures and efficiency / RNNs and normalization
M3Net: Efficient Time-Frequency Integration Network with Mirror Attention for Audio Classification on Edge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes M3Net, a ~0.03M-parameter lightweight CNN with a Mirror Attention mechanism that fuses time-flipped, frequency-flipped and center-flipped Mel-spectrogram views for edge audio classification.
Problem
High-performance audio classification models are too computationally and storage heavy for resource-constrained edge devices, so a better complexity-accuracy trade-off is needed.
Approach
Mirror Attention applies reversible spatial transformations (horizontal/time flip, vertical/frequency flip, center-symmetric flip) to Mel-spectrogram feature maps, processes each view in parallel with depth-wise separable convolutions and ECA, and multiplies the multi-view outputs (Hadamard product) to integrate local and global time-frequency features with very few parameters.
Key results
- M3Net reaches 97.44% accuracy on UrbanSound8K with 0.029M parameters, without augmentation or pre-training.
- M3Net reaches 97.03% accuracy on SpeechCommandsV2 with 0.031M parameters, statistically on par with SeqBoat and EAT-S-GMME (which use 0.293M and 1.54M parameters).
- Removing Mirror Attention drops accuracy by 22.15 points on US8K and 28.90 points on SCV2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use M3Net instead of SE-TCAM 1D CNN | Audio classification models · UrbanSound8K | accuracy (%): 94.04 -> 97.44 | Helped |
| Use M3Net instead of EAT-S-GMME | Audio classification models · SpeechCommandsV2 | accuracy (%): 97.88 -> 97.03 | Hurt |
| Use M3Net instead of SeqBoat | Audio classification models · SpeechCommandsV2 | accuracy (%): 97.35 -> 97.03 | No clear effect |
| Add Mirror Attention (ablation) | M3Net · UrbanSound8K | accuracy (%): 75.29 -> 97.44 | Helped |
| Add Mirror Attention (ablation) | M3Net · SpeechCommandsV2 | accuracy (%): 68.13 -> 97.03 | Helped |
| Retain the removed Shallow Feature Processing convolution (ablation) | M3Net · UrbanSound8K | accuracy (%): 87.31 -> 97.44 | Helped |
| Retain the removed Shallow Feature Processing convolution (ablation) | M3Net · SpeechCommandsV2 | accuracy (%): 84.26 -> 97.03 | Helped |
| Add Mirror Feature Processing (ablation) | M3Net · UrbanSound8K | accuracy (%): 87.26 -> 97.44 | Helped |
| Add Mirror Feature Processing (ablation) | M3Net · SpeechCommandsV2 | accuracy (%): 78.94 -> 97.03 | Helped |
| Add Diagonal Feature Processing (ablation) | M3Net · UrbanSound8K | accuracy (%): 93.93 -> 97.44 | Helped |
| Add Diagonal Feature Processing (ablation) | M3Net · SpeechCommandsV2 | accuracy (%): 93.08 -> 97.03 | Helped |
| Add ECA (ablation) | M3Net attention branches without transformation convolutional layers · UrbanSound8K | accuracy (%): 93.88 -> 94.56 | Helped |
Surprising findings
- Ablations show the time-flip (HMF) and frequency-flip (VMF) branches matter more for SpeechCommandsV2 than for UrbanSound8K, despite SCV2 being the simpler dataset.
- M3Net performs better on the more complex UrbanSound8K than on SpeechCommandsV2, which the authors note contradicts many previous studies.
Limitations
- Edge deployment is not tested on real devices; the authors list it as future work.
- Only two datasets, both using Mel-spectrogram input, are evaluated.
- Results rely on a 6:2:2 random split and the M3Net accuracy on SCV2 is not the highest among compared methods.
Experimental setup
Datasets: UrbanSound8K SpeechCommandsV2
Benchmarks: UrbanSound8K SpeechCommandsV2
Metrics: accuracy parameter count Wilcoxon signed-rank p-value
Baselines: AemNet-DW ULSED 2D CNN (GFCC) PhiNets M40 SE-TCAM 1D CNN DeLoRes M AdaptFormer DCLS-Delays SeqBoat EAT-S-GMME
Builds on: Efficient Channel Attention depth-wise separable convolution
Closest papers by meaning
- DTF-AT: Decoupled Time-Frequency Audio Transformer for Event Classification (AAAI 2024 · similarity 0.86)
- Audio Scanning Network: Bridging Time and Frequency Domains for Audio Classification (AAAI 2024 · similarity 0.85)
- Massive Sound Embedding Benchmark (MSEB) (NeurIPS 2025 · similarity 0.82)
- Masked Autoencoders with Multi-Window Local-Global Attention Are Better Audio Learners (ICLR 2024 · similarity 0.82)
- TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation (ICLR 2025 · similarity 0.81)
- RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation (ICLR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.