A Neuromorphic Trigger for Efficient Audio Event Detection

Technical Summary: A Neuromorphic Trigger for Efficient Audio Event Detection

Problem Statement
The increasing prevalence of transformer-based architectures and resource-intensive hardware has made the real-time processing of continuous audio streams computationally expensive and energy-inefficient. Existing Sound Event Detection (SED) and Anomalous Sound Detection (ASD) models typically process entire audio sequences, including silent or irrelevant background segments, to classify events. This "always-on" processing paradigm creates a bottleneck for edge devices and resource-constrained systems. While some efficiency gains have been attempted via binary neural networks, these often still rely on expensive temporal integration structures requiring significant memory and processing time.

Methodology
The authors propose a neuromorphic trigger system designed to act as a low-cost, class-agnostic front-end filter. The system gates input to downstream, computationally intensive classifiers, allowing them to process only salient audio segments.

• Architecture: The trigger is a lightweight, fully connected Spiking Neural Network (SNN) comprising 384 Leaky Integrate-and-Fire (LIF) neurons across four layers. It is designed to fit within TinyML standards (<1 MB memory).
• Input Processing: Audio is converted into Mel spectrograms (64 bands for URBAN-SED, 128 bands for TUT). The network outputs a binary spike train indicating the presence of an event at each timestep.
• Training: The network is trained using the Van Rossum distance as a loss function to match the output spike train against ground-truth event onsets and offsets. A surrogate gradient (Fast Sigmoid) is used to enable backpropagation through the non-differentiable spiking mechanism.
• Post-Processing: The raw spike train undergoes a "close-open" morphological filter (a concept from computer vision). This process connects discontiguous spikes (closing) and removes noise (opening) to generate contiguous blocks of "active" time.
• Pipeline Integration: The filtered output acts as a mask or switch. If the mask is active, the audio segment is passed to a downstream classifier (e.g., the Dang et al. CNN-GRU model); otherwise, the data is discarded.

Key Contributions

1. Neuromorphic Gating Mechanism: Introduction of an SNN-based trigger that selectively gates audio data, reducing the volume of information sent to heavy classifiers.
2. Class-Agnostic Detection: The trigger is trained to detect the presence of any event without classifying the specific sound type, making it applicable to both ASD and SED tasks.
3. Theoretical Cost Analysis: A framework for calculating the theoretical lower bound of computational costs (FLOPs) and energy usage when a perfect trigger is employed, demonstrating the potential for massive efficiency gains.

Results
The system was evaluated on two datasets representing different detection paradigms:

• Anomalous Sound Detection (URBAN-SED): On a class-agnostic form of the URBAN-SED dataset, the trigger achieved a segment-based F1 score of 0.97 (1-second segments). This demonstrates high reliability in distinguishing relevant audio regions from background Brownian noise. The model maintained high performance even at finer resolutions (32 ms).
• Sound Event Detection (DCASE 2017 Task 2 - TUT Rare Sounds): When combined with the Dang classifier, the trigger system showed a potential 42.6× reduction in FLOPs.
  • The trigger alone achieved a minimum event-based Acoustic Event Error Rate (AEER) of 0.416 (without post-processing) and an F1 of 0.617 (with post-processing).
  • Crucially, the authors note that if a perfect trigger were used, the downstream classifier's error rate could theoretically drop from 0.41 to 0.25. This suggests that separating the detection and classification tasks allows each model to specialize, potentially improving overall accuracy while drastically cutting costs.
• Energy Efficiency: Estimates on neuromorphic hardware (SENeCA) and standard 45 nm CMOS indicate the trigger consumes a negligible fraction of the energy required by the full system.

Significance and Claims
The paper claims that this approach offers a novel method for reducing the computational and power costs of audio processing, making it suitable for edge deployment and near-sensor computing. The authors emphasize that:

• Efficiency: The primary benefit is the substantial reduction in computational load (up to 42.6×), enabling real-time processing on constrained devices.
• Accuracy Potential: By decoupling event detection from classification, the system allows the downstream model to focus solely on classification, which the authors argue can lower error rates (as evidenced by the theoretical improvement from 0.41 to 0.25 AEER).
• Versatility: The trigger is effective for both anomaly detection in predictable environments (ASD) and event detection in noisy, complex environments (SED), though it currently performs best as an anomaly detector.

The authors conclude that while the current model is a proof-of-concept with limitations (such as sensitivity to labeling noise and the need for larger models for complex SED), it establishes a viable path toward energy-efficient, real-time audio event detection. Future work suggested includes testing on non-synthetic datasets and exploring spiking autoencoders for ASD.