Research map / Time series / Foundation models for time series
Audio Scanning Network: Bridging Time and Frequency Domains for Audio Classification
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 ASNet, a training-free audio classifier that uses echo state network reservoir readouts over MFCC, RMS and ZCR sequences, fused via a reservoir-kernel CCA, and evaluates on environmental sound and music genre classification.
Problem
Audio classification from raw audio feature sequences, where spectrogram conversion may lose temporal detail and correlations between time-domain and frequency-domain waveform fluctuations are not systematically modeled, while deep models require heavy iterative training.
Approach
Model time- and frequency-domain feature sequences (MFCC, RMS, ZCR) as readout models of echo state networks, define a reservoir kernel from L2 distances between readout models, and fuse them with a reservoir-kernel CCA (RKCCA), with no iterative training of neural weights.
Key results
- ASNet reaches 90.00% accuracy on ESC-10 with SVM, versus 81.50% for MFCC(S+D) without correlation analysis.
- ASNet reaches 92.00% accuracy on GTZAN and 82.45% on Homburg, outperforming the compared SOTA methods on those datasets.
- On UrbanSound8K ASNet achieves 92.96% accuracy, which is below AMM (93.77%) in the paper's comparison table, but its table lists AST at 90.86% and MCLNN at 74.22%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ASNet instead of AMM | Audio classification · ESC-10 | accuracy: 92.93 -> 92.96 | Helped |
| use ASNet instead of AMM | Audio classification · US8K | accuracy: 93.77 -> 91.63 | Hurt |
| use ASNet instead of Cai-FS | Audio classification · GTZAN | accuracy: 91.80 -> 92.00 | Helped |
| replace the reservoir kernel with a Gaussian kernel (ablation) | ASNet · Homburg | accuracy: 82.45 -> 93.00 | Helped |
| add dynamic audio features to static features (ablation) | Audio feature combinations · ESC-10, US8K, GTZAN, and Homburg | accuracy | Helped |
| analyze correlations between waveform fluctuations with RKCCA instead of directly merging static and dynamic features (ablation) | ASNet · ESC-10, US8K, GTZAN, and Homburg | accuracy | Helped |
| replace RKCCA with CCA (ablation) | ASNet · ESC-10, US8K, GTZAN, and Homburg | accuracy | Hurt |
| replace the reservoir kernel with a polynomial kernel (ablation) | ASNet · ESC-10, US8K, GTZAN, and Homburg | accuracy | Hurt |
| replace the reservoir kernel with a Gaussian kernel (ablation) | ASNet · ESC-10, US8K, and GTZAN | accuracy | Hurt |
| replace the reservoir kernel with a Gaussian kernel (ablation) | ASNet · Homburg | accuracy | Helped |
Surprising findings
- Linear (CCA) and polynomial kernel variants do not notably improve over simply concatenating static and dynamic features, and can underperform them.
- On UrbanSound8K ASNet is reported at 92.96 in Table 1 but the comparison table lists 91.63 for ASNet, an internal inconsistency in reported numbers.
Limitations
- Results on ESC-10 and UrbanSound8K are below or comparable to deep baselines such as AMM, and the paper claims comparability rather than superiority.
- Evaluation limited to environmental sound and music genre datasets with relatively small sizes.
Experimental setup
Datasets: ESC-10 UrbanSound8K GTZAN Homburg
Benchmarks: ESC-10 UrbanSound8K GTZAN Homburg
Metrics: accuracy F1-score
Baselines: ISHMMs MCLNN Toffa-FS AST AMM JSLRR BN Cai-FS
Builds on: Echo State Network Kernel Canonical Correlation Analysis CCA
Closest papers by meaning
- SoundCount: Sound Counting from Raw Audio with Dyadic Decomposition Neural Network (AAAI 2024 · similarity 0.86)
- M3Net: Efficient Time-Frequency Integration Network with Mirror Attention for Audio Classification on Edge (AAAI 2025 · similarity 0.85)
- DTF-AT: Decoupled Time-Frequency Audio Transformer for Event Classification (AAAI 2024 · similarity 0.84)
- Learning Temporal Resolution in Spectrogram for Audio Classification (AAAI 2024 · similarity 0.84)
- SSLAM: Enhancing Self-Supervised Models with Audio Mixtures for Polyphonic Soundscapes (ICLR 2025 · similarity 0.84)
- Synthio: Augmenting Small-Scale Audio Classification Datasets with Synthetic Data (ICLR 2025 · similarity 0.83)
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.