Research map / Architectures and efficiency / Spiking networks
Spike-based Neuromorphic Model for Sound Source Localization
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 a fully spike-driven sound source localization framework combining Resonate-and-Fire phase-locking coding (RF-PLC) for ITD encoding with a multi-auditory attention (MAA) back-end, reporting state-of-the-art accuracy and noise robustness on HRTF, Single Words, and SLoClas.
Problem
Sound source localization on edge devices needs energy-efficient spiking models that encode interaural time difference cues without costly Fourier transforms and remain robust in noisy environments.
Approach
Replace Fourier-transform-based ITD encoding with Resonate-and-Fire neurons that perform resonance filtering and phase-locked first-peak spike timing, followed by detection neurons; add a spike-driven attention module (frequency-spatial joint attention plus a short-term memory structure) that avoids MAC operations.
Key results
- On SLoClas the model reaches 0.33° MAE and 96.40% accuracy with 1.64M parameters, versus 0.60° and 95.61% for the prior best SNN (Hybrid Coding).
- Achieves 99.84% accuracy on HRTF and 99.63% on Single Words.
- MAA attention costs 9.58 uJ, about 8.49% more than the MTPC baseline (8.83 uJ), while being fully spike-driven.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RF-PLC instead of FT-based ITD extraction (ablation) | Spike-based SSL framework · SLoClas | accuracy | No clear effect |
| add ST-M (ablation) | Baseline [40] SSL model · SLoClas | MAE (◦): 1.23◦ -> 0.45◦±0.03◦ | Helped |
| add ST-M (ablation) | Baseline [40] SSL model · SLoClas | Acc (%): 93.95% -> 95.67% ±0.5% | Helped |
| add FSJA (ablation) | Baseline [40] SSL model · SLoClas | MAE (◦): 1.23◦ -> 0.49◦±0.02◦ | Helped |
| add FSJA (ablation) | Baseline [40] SSL model · SLoClas | Acc (%): 93.95% -> 95.95% ±0.6% | Helped |
| add MAA (ablation) | Baseline [40] SSL model · SLoClas | MAE (◦): 1.23◦ -> 0.33◦±0.02◦ | Helped |
| add MAA (ablation) | Baseline [40] SSL model · SLoClas | Acc (%): 93.95% -> 96.40%±0.3% | Helped |
| replace MAA with TA (ablation) | Spike-based SSL model · SLoClas | MAE (◦): 0.33◦±0.02◦ -> 0.65◦±0.05◦ | Hurt |
| replace MAA with TA (ablation) | Spike-based SSL model · SLoClas | Acc (%): 96.40%±0.3% -> 93.37% ±1.2% | Hurt |
| replace MAA with TCJA (ablation) | Spike-based SSL model · SLoClas | MAE (◦): 0.33◦±0.02◦ -> 0.47◦±0.03◦ | Hurt |
| replace MAA with TCJA (ablation) | Spike-based SSL model · SLoClas | Acc (%): 96.40%±0.3% -> 93.45% ±1.0% | Hurt |
Surprising findings
- RF-PLC ITD encoding matches Fourier-transform-based encoding accuracy with only about 1% error, despite avoiding FT.
Limitations
- No deployment on edge or neuromorphic hardware.
- Limited number of SSL datasets for validation.
- Energy figures are theoretical estimates rather than hardware measurements.
Experimental setup
Datasets: HRTF Single Words SLoClas NOISEX-92
Benchmarks: HRTF Single Words SLoClas
Metrics: MAE (degrees) accuracy (%) energy (uJ)
Baselines: MTPC-CSNN MTPC-RSNN Hybrid Coding GCC-PHAT SELDnet EINV2 SRP-DNN FN-SSL LSO MNTB MSO/LSO TA
Builds on: Multi-Tones Phase Coding (MTPC) Leaky Integrate-and-Fire neurons Resonate-and-Fire neurons Temporal attention (TA)
Closest papers by meaning
- S$^2$M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection (NeurIPS 2025 · similarity 0.87)
- Dendritic Resonate-and-Fire Neuron for Effective and Efficient Long Sequence Modeling (NeurIPS 2025 · similarity 0.85)
- FiTS: Interpretable Spiking Neurons via Frequency Selectivity and Temporal Shaping (NeurIPS 2026 · similarity 0.85)
- Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI models in Sound Localization (NeurIPS 2025 · similarity 0.85)
- Vocal Call Locator Benchmark (VCL) for localizing rodent vocalizations from multi-channel audio (NeurIPS 2024 · similarity 0.85)
- Balanced Resonate-and-Fire Neurons (ICML 2024 · similarity 0.84)
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.