Research map / Architectures and efficiency / CNN design
SoundCount: Sound Counting from Raw Audio with Dyadic Decomposition Neural Network
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 DyDecNet, an end-to-end raw-waveform network with a learnable dyadic frequency decomposition front-end and energy gain normalization, plus three polyphony-aware metrics, for counting overlapping sound events.
Problem
Counting the number of distinct, temporally overlapping (polyphonic) sound events in raw audio, which prior sound event detection and source separation methods handle poorly.
Approach
A multi-stage, coarse-to-fine dyadic decomposition of the raw waveform using learnable sinc band-pass filters (each parent filter split into approximation and detail children), combined with a learnable energy gain normalization and a density-map regression head; three polyphony-aware metrics (polyphony ratio, max polyphony, mean polyphony) quantify counting difficulty.
Key results
- DyDecNet achieves MSE 0.32 / MAE 0.73 on AudioSet vs. best baseline DPTasNet 0.81 / 1.02 (as reported in Table 1).
- On NorthEastUS, DyDecNet reaches MSE 1.01 / MAE 1.19 versus CRNNNet 1.33 / 1.77.
- Plugging the dyadic front-end into SELDNet, CRNNNet and DND-SED reduces their MSE (e.g., SELDNet 1.35 to 1.05 on NorthEastUS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DyDecNet instead of the strongest reported baseline | sound counting methods · AudioSet evaluation | MSE: 0.81 -> 0.32 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · NorthEastUS | MSE: 1.11 -> 1.01 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · Polyphony4Birds | MSE: 0.74 -> 0.46 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · Polyphony1Bird | MSE: 0.87 -> 0.54 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · TelephoneRing | MSE: 0.92 -> 0.58 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · CarEngine | MSE: 0.89 -> 0.54 | Helped |
| use DyDecNet instead of the strongest reported baseline | sound counting methods · OpenMic2018 | MSE: 0.44 -> 0.32 | Helped |
| add dyadic decomposition front-end (ablation) | SELDNet · NorthEastUS | MSE: 1.35 -> 1.05 | Helped |
| add dyadic decomposition front-end (ablation) | CRNNNet · NorthEastUS | MSE: 1.33 -> 1.20 | Helped |
| add dyadic decomposition front-end (ablation) | DND-SED · NorthEastUS | MSE: 1.19 -> 0.89 | Helped |
| replace hierarchical dyadic decomposition with single-scale filters (ablation) | DyDecNet · NorthEastUS | MSE: 0.85 -> 1.22 | Hurt |
| replace energy normalization with batch normalization (ablation) | DyDecNet · NorthEastUS | MSE: 0.85 -> 1.07 | Hurt |
Surprising findings
- Traditional signal-processing onset detectors (Librosa-onset, Aubio-onset) perform worst, with accuracy near zero on the highly polyphonic synthetic datasets.
- Source separation (DPTasNet) performs worse than SED-based methods on high-polyphony bioacoustic data.
Limitations
- Real-world NorthEastUS performance remains far lower than synthetic datasets, indicating difficulty on real recordings.
- Model capacity and training data may need to scale to handle much larger bird class diversity, per the authors.
Experimental setup
Datasets: AudioSet NorthEastUS Polyphony4Birds Polyphony1Bird TelephoneRing CarEngine OpenMIC TAU-NIGENS
Benchmarks: AudioSet NorthEastUS Polyphony4Birds Polyphony1Bird TelephoneRing CarEngine OpenMIC TAU-NIGENS
Metrics: MSE MAE accuracy rate polyphony ratio max polyphony mean polyphony ER F-score LE LR
Baselines: Librosa-onset Aubio-onset SELDNet CRNNNet DND-SED DPTasNet
Builds on: SincNet SELDNet CRNN DPTasNet Librosa
Closest papers by meaning
- Audio Scanning Network: Bridging Time and Frequency Domains for Audio Classification (AAAI 2024 · similarity 0.86)
- SALMONN: Towards Generic Hearing Abilities for Large Language Models (ICLR 2024 · similarity 0.84)
- AND: Audio Network Dissection for Interpreting Deep Acoustic Models (ICML 2024 · similarity 0.83)
- FlowDec: A flow-based full-band general audio codec with high perceptual quality (ICLR 2025 · similarity 0.83)
- DeepASA: An Object-Oriented Multi-Purpose Network for Auditory Scene Analysis (NeurIPS 2025 · similarity 0.83)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.82)
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.