Paper deep dive
A Neuromorphic Trigger for Efficient Audio Event Detection
Benjamin Hatton, Oliver Rhodes, Luca Peres
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 6/21/2026, 1:06:16 AM
Summary
This paper proposes a neuromorphic trigger for efficient audio event detection using a lightweight Spiking Neural Network (SNN). The trigger acts as a low-cost front-end that identifies salient audio segments and selectively gates input to a more computationally intensive downstream classifier. The system was evaluated on two tasks: Anomalous Sound Detection (ASD) using the URBAN-SED dataset, achieving a one-second segment-based F1 score of 0.97, and Sound Event Detection (SED) using the DCASE 2017 Challenge Task 2 dataset. When combined with the Dang classifier, the trigger achieved a 42.6x reduction in FLOPs while reducing the event-based error rate from 0.41 to 0.25, demonstrating its potential for real-time, energy-efficient edge computing.
Entities (9)
Relation Signals (5)
Neuromorphic Trigger → evaluatedon → URBAN-SED
confidence 100% · For ASD, the trigger achieves a one-second segment-based F1 score of 0.97 on a class-agnostic form of the URBAN-SED dataset
Neuromorphic Trigger → evaluatedon → DCASE 2017 Challenge Task 2
confidence 100% · For SED, the trigger is combined with the Dang classifier on the DCASE 2017 Challenge Task 2 dataset
Neuromorphic Trigger → gatesinputto → Dang Classifier
confidence 100% · The output of this filter triggers a switch that allows audio to be passed to a larger classifier.
Neuromorphic Trigger → implementedas → Spiking Neural Network
confidence 100% · The proposed trigger is implemented as a lightweight fully connected SNN
Leaky Integrate-and-Fire → isatypeof → Spiking Neural Network
confidence 100% · the Leaky Integrate and Fire (LIF) model is a spiking neuron
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Efficient processing of continuous audio streams remains a key challenge for real-time and resource-constrained systems. This paper introduces a neuromorphic trigger for audio event detection, based on a spiking neural network (SNN) that selectively gates input to downstream models. The proposed trigger acts as a low-cost front-end, identifying salient audio segments and forwarding only these to a more computationally intensive model for tasks such as classification. The trigger is implemented as a lightweight fully connected SNN and evaluated on two representative tasks: Anomalous Sound Detection (ASD) and Sound Event Detection (SED). For ASD, the trigger achieves a one-second segment-based F1 score of 0.97 on a class-agnostic form of the URBAN-SED dataset, demonstrating high reliability in identifying relevant audio regions. For SED, the trigger is combined with the Dang classifier on the DCASE 2017 Challenge Task 2 dataset, showing a potential $42.6\times$ reduction in FLOPs while reducing the lower bound of the event-based error rate from 0.41 to 0.25. These results highlight the potential of neuromorphic triggers as real-time, energy-efficient front-end filters, enabling substantial reductions in computational cost.
Tags
Links
- Source: https://arxiv.org/abs/2606.17775v1
- Canonical: https://arxiv.org/abs/2606.17775v1
Trouble viewing inline? Open PDF directly →
Full Text
49,792 characters extracted from source content.
Expand or collapse full text
A Neuromorphic Trigger for Efficient Audio Event Detection Benjamin Hatton 0009-0007-5509-7873 ICNS, University of ManchesterManchesterUK , Oliver Rhodes 0000-0003-1728-2828 ICNS, University of ManchesterManchesterUK and Luca Peres 0000-0001-9748-9073 ICNS, University of ManchesterManchesterUK (5 June 2009) Abstract. Efficient processing of continuous audio streams remains a key challenge for real-time and resource-constrained systems. This paper introduces a neuromorphic trigger for audio event detection, based on a spiking neural network (SNN) that selectively gates input to downstream models. The proposed trigger acts as a low-cost front-end, identifying salient audio segments and forwarding only these to a more computationally intensive model for tasks such as classification. The trigger is implemented as a lightweight fully connected SNN and evaluated on two representative tasks: Anomalous Sound Detection (ASD) and Sound Event Detection (SED). For ASD, the trigger achieves a one-second segment-based F1 score of 0.97 on a class-agnostic form of the URBAN-SED dataset, demonstrating high reliability in identifying relevant audio regions. For SED, the trigger is combined with the Dang classifier on the DCASE 2017 Challenge Task 2 dataset, showing a potential 42.6×42.6× reduction in FLOPs while reducing the lower bound of the event-based error rate from 0.41 to 0.25. These results highlight the potential of neuromorphic triggers as real-time, energy-efficient front-end filters, enabling substantial reductions in computational cost. Spiking, Neuromorphic, SNN, Anomaly Detection, Audio, Edge Computing, Near-Sensor, SED †copyright: acmlicensed†journalyear: 2026†doi: X.X†conference: Make sure to enter the correct conference title from your rights confirmation email; August 04–06, 2026; Chicago, IL†isbn: 978-1-4503-X-X/2018/06†ccs: Computing methodologies Neural networks†ccs: Computing methodologies Bio-inspired approaches†ccs: Computing methodologies Feature selection†ccs: Computing methodologies Temporal reasoning Figure 1. Proposed pipeline for more efficient processing of data. The audio is converted into a Mel spectrogram that is then passed to a spiking network (Spiking Event Detector). These spikes are then passed through a close-open filter to create contiguous blocks. The output of this filter triggers a switch that allows audio to be passed to a larger classifier. Shows raw audio being processed into a Mel spectrogram, followed by a spiking network triggering a switch for the original audio to be passed to the large classifier. 1. Introduction In a world where AI usage is ever expanding, methods of reducing the computational and power costs of such systems are growing in demand. Modern-day data processing is growing far more expensive than it has been in the past (Sevilla et al., 2022). This is due to the widening prevalence of transformer-based architectures, as well as the more intensive hardware required to run such models. With this in mind, recent research has attempted to find ways to circumvent these increased costs and various approaches have been taken; for example, binary neural networks, spiking neural networks, and dynamic networks have all attempted to simplify the computations performed during inference. This work applies such principles to audio event detection. In this paper, audio event detection is used as an umbrella term to encompass two similar but distinct tasks: sound event detection (SED) and anomalous sound detection (ASD). Sound event detection is the problem of detecting and classifying certain patterns within an audio sequence, known as events (Stowell et al., 2015). This can be viewed as events happening in environmental audio. For example, in the DCASE 2017 Challenge (Task 2), the task involved identifying gunshots, glass breaks, and baby cries in a stream of background noise (Mesaros et al., 2017). Most existing models for SED are designed to classify audio sequences into event classes while simultaneously identifying the on- and offset of the event. This combination of tasks forces the whole sample to be processed in its entirety, including the time between events of interest. This greatly increases the processing requirements compared to a model in which the data is filtered first. The majority of existing architectures for sound event detection follow the same rough archetype. For example, convolutional recurrent neural networks (CRNNs) can be used for the detection of events in domestic environments (De Benito-Gorrón et al., 2021). Alongside these traditional networks, gated recurrent units (GRUs) and long short-term memory (LSTM) layers are often used (Lu, 2017; Adavanne et al., 2017). For more difficult tasks within SED, e.g. using heterogeneous data with missing labels, transformers are currently the state of the art (Schmid et al., 2024). Anomalous sound detection (ASD) is the problem of determining whether some sound emitted from an object is normal or anomalous (Wang et al., 2021). The problem does not concern the classification of the sound, nor the identification of the cause. The state of the art for this problem is a mix between audio transformers and autoencoders (Wang, 2025; Yang, 2025). Despite some attempts to reduce the power consumption of modern sound event and anomalous sound detection systems, e.g. with binary neural networks (Cerutti et al., 2020), these networks still need to use expensive structures that integrate the data over time, requiring large amounts of memory and processing time. As spiking neural networks (SNNs) seamlessly integrate over time, they are naturally applicable to tasks with sequential data (Maass, 1997). This makes them a suitable choice for audio-based tasks such as speech recognition (Cramer et al., 2022). Additionally, spiking networks have shown improvements to power costs when applied to other auditory tasks, e.g. keyword spotting (Blouw et al., 2019). Due to this — and the pervasiveness of cloud computation — a logical step toward efficiency is to reduce the amount of data being processed. To capitalise on this, we propose an SNN trigger model to be placed before a large classifier (as in Figure 1). This model reads the audio input and detects sound events therein but does not attempt to classify them. The output represents a mask for the audio sequence, where a negative value for the mask means that the audio represented will not be processed. This greatly reduces the required processing, as only the salient parts will be analysed. 2. Related Works 2.1. Spiking Neural Networks Spiking neural networks (SNNs) are often called the third generation of neural networks (Maass, 1997). They are designed to function more like the brain, taking great inspiration from biology, partially by introducing a temporal dimension to most problems. Rather than all neurons outputting a real number at every step, the spiking neurons emit a binary spike depending on prior inputs. The most consequential part of this is the inherent integration over time apparent in many spiking neuron models. For example, the Leaky Integrate and Fire (LIF) model is a spiking neuron that consists of a membrane potential and an output (Abbott, 1999). The original work details a continuous-time mathematical model; however, for a digital system, a discrete-time version is typically used. The snnTorch implementation of the LIF neuron is one such version (Eshraghian et al., 2023). For a given neuron, its membrane potential is represented by U[t]U[t], and its output by S[t]S[t] for a timestep t. For a neuron to spike, U must exceed the spiking threshold for that neuron, UthrU_thr. Therefore, at a given time t, we obtain the Heaviside step function: (1) S[t]=1U[t−1]>Uthr0OtherwiseS[t]= cases1&U[t-1]>U_thr\\ 0&Otherwise cases The reset mechanism is how the membrane potential changes after a spike is emitted. The method used in this work is subtraction, which can be mathematically defined as: (2) U[t]=βU[t−1]+Iin[t]−UthrS[t]=1βU[t−1]+Iin[t]OtherwiseU[t]= casesβ U[t-1]+I_in[t]-U_thr&S[t]=1\\ β U[t-1]+I_in[t]&Otherwise cases Iin[t]I_in[t] represents the input to the neuron at time step t. The value β is the decay rate of the neuron, hence its “leaky” nature. The decay dictates the “memory” of the neuron, with higher values retaining more information over longer periods of time. As the LIF functionality hinges on a Heaviside step function, the spike generation is non-differentiable. Therefore, to train via backpropagation, a surrogate gradient is typically used (Neftci et al., 2019). Surrogate gradients approximate the gradient of a neuron during a backward pass by replacing the Heaviside function with a smooth gradient function, e.g. fast sigmoid and atan, enabling credit assignment for multilayered networks. 2.2. Audio Event Detection Evaluation Within audio event detection, there are two main evaluation paradigms used to assess performance: event-based and segment-based (Mesaros et al., 2016). The event-based paradigm acts as a holistic approach, treating the predictions as block events. To evaluate the timing of events, there is often a margin (“collar”) in which the model can err without being penalised. Standards for the collar vary depending on the dataset, with 500 ms being common for the tougher DCASE challenges (Mesaros et al., 2017). The two paradigms are most commonly evaluated using the following metrics: the F1 score and the acoustic event error rate (AEER). Both of which are calculated by counting the true positives (TPs), false positives (FPs), and false negatives (FNs). For event-based metrics these are counted as: • True Positive: the model output has an overlapping event with the same label as the ground truth. • False Positive: the model output has an event in the same time frame but no correlation with the label. • False Negative: an event in the reference has no correlation with an event in the system’s output. F1 score is the harmonic mean of precision (the percentage of positive outputs that are true), P, and recall (the percentage of correct answers that are recognised), R. An F1 score of 11 would indicate perfect recognition. (3) P=TPTP+FP,R=TPTP+FN,F1=2PRP+R.P= TPTP+FP, R= TPTP+FN, F_1= 2PRP+R. Acoustic event error rate (AEER or error rate) is calculated by summing the substitutions (S), deletions (D), and insertions (I) and then calculating the mean over the number of events in the sample (N) (as shown in Eqs. 7). Each of the metrics can be calculated using just false positives (FP) and false negatives (FN). A perfect model would achieve an AEER of 0. For any given sample k∈Kk∈ K: (4) S(k)=min(FN(k),FP(k))S(k)=min(FN(k),FP(k)) (5) D(k)=max(0,FN(k)−FP(k))D(k)=max(0,FN(k)-FP(k)) (6) I(k)=max(0,FP(k)−FN(k))I(k)=max(0,FP(k)-FN(k)) (7) AEER AEER =∑kK[S(k)+D(k)+I(k)]∑kKN(k) = Σ^K_k [S(k)+D(k)+I(k) ]Σ^K_kN(k) (8) =∑kKmax(FN(k),FP(k))∑kKN(k) = Σ^K_kmax (FN(k),FP(k) )Σ^K_kN(k) The segment-based paradigm is similar, except it divides the sample into adjacent sequences of length n. If there is any event in the output, then that segment is labelled as positive by the model. This is then compared to a similarly divided ground truth to determine the above statistics. 2.3. DCASE 2017 Task 2 The DCASE Challenge 2017 Task 2 is chosen as one option to explore the neuromorphic trigger system due to its sparse event data and real-world similarities (Mesaros et al., 2017). The goal of the challenge is to identify the onset, offset, and class of events occurring within a sequence. The dataset used is the TUT Rare Sounds 2017 dataset, which is a set of monophonic audio event samples. The challenge solutions vary in size, method, and efficacy (Mesaros et al., 2019). These are all then evaluated on a separate set from the dev-train and dev-test sets, resulting in two event-based metrics: F1 and error rate. For ranking the solutions, AEER is used over F1. Figure 2 shows the AEER of the network-based solutions against computational cost. As can be seen in the figure, the method from Lim et al. (2017) performs best with an average computational cost, using a unidirectional LSTM with a convolutional backbone. This is then followed by Cakir and Virtanen (2017) using a CRNN with a good AEER score of 0.1733 but a high FLOP count. The highest cost among the models is Dang et al. (2017) despite a relatively low AEER of 0.4107. Within the challenge, many different approaches were employed. Some, like Phan et al. (2017) attempt to teach background reduction as well as classification; while others, like Cakir and Virtanen (2017), use ensembles of the models that have the same structure but different training. These approaches varied in terms of architecture, including convolutional networks (CNNs) (Dang et al., 2017), autoencoders (Ravichandran and Das, 2017), long short-term memory cells (LSTMs) (Lim et al., 2017; Li and Li, 2017), and gated recurrent units (GRUs) (Wang et al., 2017). Figure 2. Comparison of the error rate (AEER) and the number of FLOPs of DCASE 2017 Challenge entries. The number of FLOPs is calculated using calflops (xiaoju ye, 2023), by passing through a 128 timestep sample, then using that to calculate FLOPs over the entire dataset. Also included is the trigger model introduced in Section 3.2 and evaluated in Section 4.1.2. 3. Methods 3.1. Datasets Within audio event detection, there are two forms of dataset: polyphonic and monophonic. A polyphonic set contains audio in which events are layered over background noise and possibly over other events as well. A monophonic set never overlaps two event sounds, creating a simpler classification task. Two datasets were used to evaluate the efficacy of the model. The first one, URBAN-SED, is a synthetic dataset generated using Scaper for polyphonic SED (Salamon et al., 2014, 2017). The dataset consists of Brownian noise layered with various sound events, such that some sounds may occur simultaneously with others. URBAN-SED has ten classes, such as a dog bark or a car horn. The difficulty within this dataset is primarily in polyphonic recognition rather than event detection itself, due to the consistency of background Brownian noise. The second dataset is from the DCASE 2017 Challenge for Task 2 (Mesaros et al., 2017), the TUT Rare Sound Events 2017. This set is far more complex than URBAN-SED, despite being both monophonic and synthetic. The goal of the challenge was to generate datasets with real-world recordings; i.e. noise was recorded at various locations and overlaid with the events to be identified. The way in which the clips are overlaid often means that the events are very quiet and sunk into the background. This makes them difficult to identify, even for the human ear. On top of this, only 50%50\% of the samples contain an event, so no assumption about presence can be made; however, there is only ever at most one event within a sample. The two datasets represent two slightly different tasks in audio event detection. The TUT dataset, which has a noisy and variable background, requires sifting through the noise to try to find the recognisable sounds and is a prime example of SED. This could be applied to keyword spotting (López-Espejo et al., 2022) or identifying alarming noises in urban environments. On the other hand, URBAN-SED, due to its consistent background noise, represents the identification of abnormalities occurring in a relatively quiet, clean, and predictable environment. Therefore, it is a task far closer to anomalous sound detection (ASD) (Kawaguchi and Endo, 2017) due to the trigger being class-agnostic. This is an application more akin to wildlife monitoring or machine malfunction detection. 3.2. Near-Sensor Trigger The proposed solution comprises a 4-layer SNN made up of 384 LIF neurons (full structure and parameters in Table 1) and is situated at point (a) in the pipeline (Figure 1). This network is structured with a single output representing whether further processing is needed at a given timestep. The size of 384 neurons was selected to demonstrate the small scale possible, fitting well within the TinyML standard of less than 1 MB memory (Heydari and Mahmoud, 2025). An SNN is used as it is low-power and low-latency when combined with neuromorphic hardware, further reducing computational costs. The output is used as a trigger for a large classifier, as seen in Figure 1. The input of the trigger is a Mel spectrogram using 128 mel-bands for the TUT dataset and 64 mel-bands for the URBAN set. This spectrogram is created with a window of 64 ms and a hop of 32 ms, allowing for a 50% overlap at every timestep. It was trained using the Adam optimiser with a learning rate of 0.0010.001 and a scheduler to reduce it on plateau (with a patience of 1515 epochs). Layer Dimension 1 64/128→12864/128→ 128 2 128→128128→ 128 3 128→127128→ 127 4 127→1127→ 1 LIF Parameter Value Decay (β) 0.8 Reset Mechanism Subtraction Surrogate Gradient Fast Sigmoid Threshold (UthrU_thr) 1.01.0 Table 1. The structure and parameters of the proposed trigger network. The dimensions on the left detail the fully connected layers of LIF neurons. A table displaying each of the 4 layers of the near-sensor spiking trigger network. Each layer is a set of fully-connected LIF neurons. To train this model, a target output spike train is used. The datasets used supply on- and off-set times for each of the events. All timesteps situated between these two times are then labelled as positive. When comparing the output of the network to the target spike train, the Van Rossum distance is calculated (Van Rossum, 2001). This is then used as the loss function for training the network through backpropagation, as in (Zhang and Li, 2020; Yang et al., 2021). The output spike train of the model is then post-processed using a close-open filter (situated at (b) in Figure 1), a concept from computer vision (Serra, 1982). Opening and closing are operations comprising erosion and dilation (Matheron, 1975). Each of these is used on binary images to shrink and expand the size of the positive values within the image. Modelling the spike train as a one-dimensional binary image allows for the application of these processes, as seen in Figure 3. The post-processing more effectively defines the mask as it connects groups of discontiguous spikes (closing) and then removes any noise (opening). Any set of spikes that persists through the opening operation is seen as significant enough to be processed by the large classifier. The parameter “expansion” dictates the size of the filter, e.g. Figure 3 shows an expansion of 22. Figure 3. A diagram of the closing then the opening of a spike train modelled as a binary image. Here, white/black squares represent the presence/absence of spikes at a given timestep. A diagram showing the closing then opening of the same spike train (modelled as a binary image) with each individual step labelled. 3.3. Post-Trigger Classifier As almost all models for this task classify the events, a post-trigger classifier must be used to enable direct comparison (situated at (c) in Figure 1). To do so, the TUT dataset was chosen as it has the most results available. To show improvement, a model from the DCASE challenge was selected. Dang et al. (2017) created a CNN- and GRU-based classifier with performance captured in Figure 2. This model was chosen due to its performance and because it has the most comprehensive technical report, enabling reproduction. The Dang solution consists of three separate models, each acting as a binary classifier between its class and the other two. Within each model are two parallel pipelines, one comprising a CNN stack and the other a GRU. The two outputs are then aggregated and processed through a fully-connected (FC) layer to produce the model output. The three results are then concatenated and returned, with final prediction processing depending on the task. The model is trained through backpropagation in two separate ways. One version treats each of the individual pipelines as their own model and uses binary cross entropy loss. The other version treats it as a complete network, using cross entropy loss (XE). As this is a conventional neural network, no surrogate gradients are needed. A standard Adam optimiser is used with a learning rate of 0.0010.001, and a scheduler that reduces the learning rate on plateau. As preprocessing, the data is cut to 100-timestep event samples using the supplied onset and offset to emulate the effect of the trigger. If the selected region is less than 100 timesteps, the data is padded on both sides with data taken from the initial sample. If the event is longer than 100 timesteps, it is truncated. The result of the model is a 3-tuple representing each class. As part of post-processing, the results are passed through a softmax function to turn the logits into a probability distribution (Bishop, 2006; Goodfellow et al., 2016). After this, the 3-tuple is passed through a single FC layer to calculate a learned weighted sum of the output. The FC layer consists of three input neurons and three output neurons, acting as a linear transformation of the outputs to optimise the selection process. 3.4. Evaluation The metrics used to quantify performance depend on the dataset. For the URBAN-SED dataset, one second segment-based scores are used as a standard, as seen in (Koga et al., 2024; Venkatesh et al., 2022). The performance is measured using the F1 score, as this is the standard for ASD. All of the models found in the literature for this dataset predict class as well as onset and offset, e.g. the transformer-based architecture in (Ye et al., 2021). As TUT Rare Sounds 2017 is a challenge dataset, there are several models to compare with. Evaluation for this dataset is primarily done using error rate rather than F1 as a standard, showing that mistakes are penalised more than correct predictions. We follow this trend as it allows for greater comparability between models. Again, the comparison models are classifiers that also predict onset and offset, rather than the separation of the tasks attempted here. When evaluating the TUT set, the fact that there is at most one event per sample allows for the simplification of post-processing. After applying the close-open filter, the largest contiguous block of spikes can be taken from the output spike train to provide the prediction for the sample. Besides the statistical evaluation metrics, the number of Floating Point Operations (FLOPs) is counted and compared to those of other models. This is to show the reduction in processing requirements by using the trigger. 3.4.1. Evaluation Challenges For a class-agnostic trigger, there are some challenges in terms of evaluation. Firstly, no information on similar models could be found; therefore, all evaluations must be done through comparisons to classifying models. This causes a discrepancy in the results, as the trigger does not focus on the classification of sounds but simply on detecting their presence. As well as this, a spiking trigger can be incredibly sensitive. This is compounded by mistakes in the labelling of datasets. Koga et al have found that various annotators often label the data extremely differently, and this disparity can greatly affect the performance of detection models by upwards of 30%30\% on URBAN-SED (Koga et al., 2024). 4. Results 4.1. Near-Sensor Trigger 4.1.1. URBAN-SED The model, when trained on URBAN-SED, achieved a segment-based F1 score of 0.9724 using the standard one second segment length. This shows that the trigger can effectively detect almost any event within the test set while not misidentifying the background noise. A comparison of various expansions and segment lengths can be seen in Table 2. Within the table, it is clear that expansion generally does not affect the results, except in two scenarios. First, having any expansion effectively removes errant spikes, as an expansion of zero allows noisy spikes to remain, decreasing the F1 score compared to other expansions. The second scenario is when the expansions almost exceed, or exceed, the length of the segment being analysed. This allows the output spike train to become more consistent by bridging gaps and, therefore, bridging the segments that would otherwise be empty. The reason for the decrease in F1 at higher expansions in Table 2 is generally due to connections being made, sometimes spanning segments where they should not be. This causes a surge in false positives (reducing precision), dropping the F1 score. These surges are not as impactful as they could be because the higher expansion also decreases false negatives (increasing recall), mostly mitigating the drop in F1. Therefore, an expansion should be selected based on whether recall or precision is more important for the application. The reduction in F1 as the segment resolution gets finer is expected, as the small discrepancies between the target and predicted spike trains become more prominent. The difference between the one second segment and the 32 ms segment scores is low at 0.0215, showing that the model performs well at a range of resolutions with consistent output. The models trained for SED on this dataset achieve a range of results. The You Only Hear Once model, based on the YOLO paradigm, achieves an average one second segment-based F1 score of 0.59 (Venkatesh et al., 2022). The transformer-based model achieves a score of 0.6577 (Ye et al., 2021) and a separate unspecified model achieves a score of 0.7276 (Koga et al., 2024). Despite the model proposed here attaining a higher score, it is not truly comparable due to these results including classification. Exp. Seg. Len. 1 s 0.5 s 0.25 s 0.1 s 32 ms 0 0.9718 0.9661 0.9590 0.9504 0.9455 1 (32 ms) !150.9724 !150.9668 !150.9592 0.9505 0.9500 2 (64 ms) 0.9718 0.9659 0.9588 0.9503 0.9505 3 (96 ms) 0.9713 0.9656 0.9588 !150.9506 0.9508 4 (128 ms) 0.9708 0.9656 0.9589 !150.9506 !150.9509 5 (160 ms) 0.9705 0.9655 0.9589 !150.9506 !150.9509 6 (192 ms) 0.9702 0.9656 0.9591 !150.9506 0.9508 7 (224 ms) 0.9696 0.9654 0.9590 0.9503 0.9504 8 (256 ms) 0.9686 0.9650 0.9585 0.9496 0.9497 9 (288 ms) 0.9682 0.9648 0.9580 0.9489 0.9489 Table 2. Comparison of the segment-based F1 scores of the URBAN-SED tests using various expansions and segment lengths. Note that 32 ms is the length of one frame so represents a frame-level score. 4.1.2. TUT Rare Sounds 2017 For TUT Rare Sounds, the lowest event-based AEER achieved by the proposed model is 0.416 (with an F1 of 0.5840.584). This was attained without using the close-open filter. The best F1 attained was 0.6170.617 (with an AEER of 0.4730.473) using an expansion of 5 for the close-open filter. Exp. AEER F1 Prec. Rec. TPs FPs FNs 0 !150.416 0.5940.594 0.605 !150.584 438 286 312 1 0.4320.432 0.6110.611 0.661 0.568 426 218 324 2 0.4410.441 0.6110.611 0.675 0.558 419 201 331 3 0.4480.448 0.6110.611 0.685 0.552 414 190 336 4 0.4570.457 0.612 0.701 0.542 407 173 343 5 0.4730.473 !150.617 0.746 0.526 395 134 355 6 0.4860.486 !150.617 !150.774 0.513 385 112 365 7 0.5090.509 0.5960.596 0.761 0.490 368 115 382 8 0.5290.529 0.5780.578 0.749 0.470 353 118 397 9 0.5620.562 0.5510.551 0.745 0.437 328 112 422 Table 3. Results of the Trigger on the TUT Rare Sounds 2017 dataset with various levels of expansion. These comprise the Error Rate (AEER), Precision, Recall, True Positives (TPs), False Positives (FPs), and False Negatives (FNs). The full set of results for the different expansions can be found in Table 3. Figure 2 shows the performance of the trigger concerning the FLOP count of each of the other models tested. It must be noted, however, that this is solely the detection of an event rather than the classification of the sample. The performance gained and lost through the close-open filter occurs due to the nature of the two metrics. Due to the simplification of the error rate in Eq. 88, it can be seen that an increase in expansion increases only one of these statistics (in this case, it is FNs, as seen in Table 3). Comparatively, the F1 score is based on a ratio between precision and recall, and so a range of statistics affect it. It can be seen that the lower expansions are generally best for recall, as what could be mistaken for erroneous spikes is not removed. The higher expansions achieve far greater precision but begin to miss the more obscure events. Table 4 shows the class-wise detection results of the network. It seems that the length of each sound greatly alters the performance of the filter on the resultant trains. As can be seen, the “glass break” class (an average of 0.790.79 seconds long) is the easiest to detect for all tested expansions except for the largest two. The “gunshot” (0.860.86 s) is the hardest to detect and is most easily identified with no expansion. The “baby crying” (1.821.82 s) is difficult, but some improvement is shown at higher expansions as expected. This is due to the spike trains closing the gap between wails. The gunshot and the glass breaking perform better at lower expansion as they are shorter, sharper sounds than the baby. Seemingly, the longer the event being identified, the better the expansion performs; however, more event classes need to be analysed to draw a stronger conclusion. AEER F1 Exp. Baby Glass Gun Baby Glass Gun 0 0.504 0.364 !150.500 0.524 0.713 !150.530 1 0.492 !150.276 0.552 0.558 !150.739 0.517 2 0.480 !150.276 0.568 0.575 0.731 0.510 3 0.480 0.296 0.568 0.584 0.721 0.513 4 0.464 0.328 0.580 0.604 0.707 0.510 5 0.460 0.352 0.608 0.627 0.709 0.500 6 0.464 0.372 0.624 !150.636 0.708 0.490 7 0.464 0.408 0.656 0.627 0.685 0.459 8 !150.456 0.448 0.684 0.634 0.650 0.429 9 0.460 0.504 0.724 0.635 0.606 0.387 Table 4. A table showing the results of the trigger system on the individual classes of TUT Rare Sounds 2017. The AEER and F1 are measured at various expansions. 4.1.3. Theoretical Costs Assuming a perfect trigger system, theoretical computational costs can be calculated. The number of FLOPs that the trigger uses over the entire evaluation set is FtriggerF_trigger. An “event mask”, E(t)E(t), represents the binary presence of an event at timestep t∈Tt∈ T, i.e. if an event is present at t then the function returns 11; otherwise, it returns 0. For a given classifier, the number of FLOPs it uses per timestep is represented by the constant C giving the equation: (9) FLOPs=Ftrigger+∑tT(E(t)⋅C)FLOPs=F_trigger+Σ^T_t (E(t)· C ) Using this, the theoretical cost lower bound can be found for each classifier, as shown in Figure 4. As can be seen, the potential improvement is inversely proportional to the number of sound events in the stream, with larger classifiers using orders of magnitude fewer FLOPs for inference. Alongside FLOPs, an estimate of the energy was calculated (Table 5). This was done using two different sets of hardware to map the spiking and non-spiking operations within the model. SENeCA was used as the target neuromorphic platform (Tang et al., 2023), with multiply-accumulate (MAC) operations theoretically calculated from the given instruction-level data. For SENeCA, synaptic operations (SOps) were counted through accumulates (ACs) as a vector of binary spikes does not require multiplication (Lemaire et al., 2022). Aside from this, the arithmetic cost on 4545 nm hardware was calculated using the numbers from Han et al. (2015) as cited in the literature (Kim et al., 2020; Ji et al., 2023; Wu et al., 2024). For both forms of hardware, the energy cost for each operation type is calculated as: (10) Eusage=NopEopE_usage=N_opE_op Where NopN_op is the number of operations and EopE_op is the energy cost of said operation. While these number would require real-world testing and validation, they demonstrate the small fraction of energy consumed by the trigger relative to the system-level consumption. Figure 4. A comparison of the theoretical FLOP count assuming a perfect trigger system, calculated as in Section 4.1.3, with the original model. The grey lines represent the shift of each model, with the dotted black line representing the FLOP count of the trigger itself. Hardware Op. Num. pJ/Op Energy HW Total SENeCA(Tang et al., 2023) SOp. 6.56.5M 13.313.3 87.3 µJ87.3\, J 359.28 µJ359.28\, J MAC 15.415.4 M 17.717.7 272 µJ272\, J 45 nm(Han et al., 2015) 4.64.6 70.7 µJ70.7\, J 76.599 µJ76.599\, J AC 6.56.5M 0.90.9 5.91 µJ5.91\, J Table 5. A table showing the energy estimates of running the trigger network on two different forms of hardware, SENeCA and 45 nm CMOS (Tang et al., 2023; Han et al., 2015). This is separated into the respective operations: synaptic operations (SOp), multiply-accumulates (MACs), and accumulates (ACs). Each of the figures is an average for each 30-second sample from the TUT 2017 dataset. 4.1.4. Comparison As stated in Section 3.1, the two datasets represent different tasks. The URBAN-SED set, with its background of Brownian noise, represents an analysis of anomalies layered in predictable noise (ASD). From the results, it can be seen that the proposed solution is effective for this style of problem. In the TUT dataset, the trigger is effective for its size, but the room for improvement is clear. Given a larger model, it is likely that the trigger can learn to better spot the more hidden sounds. The trade-off, however, is that a larger size increases latency and computational costs. From the data collected, it can be seen that the trigger is currently most effective for anomaly detection rather than sound events. Despite this, the results show great potential for SED in terms of reduction in computational cost. 4.2. Post-Trigger Classifier Table 6 shows the accuracy achieved by the classifier for various training methods. Note that the accuracy is calculated only with samples in which an event occurs. As can be seen, the binary classifiers achieve accuracy above 90%90\%, but when they are combined into a 3-class classification, the accuracy falls. There is negligible difference in accuracy when comparing the complete model with and without an FC layer attached. The AEER and F1 of the two models also differ insignificantly. When comparing the XE trained model, the parallel trained models perform better with a five percentage point increase in accuracy. Despite this, the XE model still performs relatively similarly to the others in terms of AEER, likely due to the trigger system’s performance on the dataset. The entire pipeline performs worse on the task than most of the other models listed in the challenge (performing better than the Li and baseline models). This is due to the combination of inaccuracy in the model and the error rate of the trigger. In fact, using a perfect trigger allows the binary trained model without an output layer to achieve an error rate of 0.2573, which is better than the original Dang model (which attained 0.4107). Using this theoretical trigger also drastically decreases the FLOP count by 42.6×42.6×. Not only can the trigger system increase the accuracy of a model in such a problem, but it can also greatly reduce the computational cost. This shows that the separation of SED into sub-problems simplifies the task, allowing the individual parts to be solved more effectively. Method Section Accuracy AEER F1 Baby 92.73%92.73\% !40 !40 Parallel Glass 90.60%90.60\% !40 !40 Binary Gunshot 93.27%93.27\% !40 !40 Training Total (FC) 79.60%79.60\% 0.608 !150.454 Total (Sep.) 79.73%79.73\% !150.596 0.433 XE Loss Total 74.80%74.80\% 0.620 0.521 Table 6. Post-trigger classifier results. Binary accuracies use the previously mentioned 100-timestep samples that contain events. AEER and F1 represent TUT Rare Sounds 2017 scores using trigger outputs. 5. Conclusion In this work, we present a method that can greatly reduce the computational costs of analysing audio through the use of a trigger. Through theoretical FLOP analysis, it is clear that any form of pre-classifier trigger significantly reduces the computational cost of inference, achieving up to 42.6×42.6× for the tested Dang model. As well as this, the “ideal” trigger displays the potential improvement that can be gained in terms of accuracy for classifier models, with the error rate of the classifier dropping from 0.41 to 0.25. The separation of tasks within sound event detection allows each model to focus more on its individual problem, decreasing the error rate. Alongside this, it is evident that for anomalous sound detection, small trigger systems are effective at removing a vast amount of unnecessary data from processing. The 0.97 F1 score achieved displays its ability to select salient sections for further analysis while rarely including background sections. This is further supported by the performance at smaller segment lengths than standard (e.g. the one frame size of 32 ms), indicating that the trigger operates effectively at high resolutions. This initial investigation shows promising results as well as limitations worth exploring. First, the applicability of the model can be tested beyond synthetic datasets to further validate the presented approach. Alternatively, an exploration of keyword spotting and human speech detection using the trigger could prove its viability in low-power accessibility systems. Secondly, the current model focussed on a single structure and size, but different styles of triggers may also prove effective. Traditionally, ASD solutions are based on autoencoders. These can be created for a variety of scales and have had success with spiking neurons. Therefore, an analysis of small-scale spiking autoencoders may show improvements in computational cost in ASD. Finally, while this work shows potential for the simplified problem of classification after the identification of an event (in Section 4.2), further research into the topic is warranted. A potential avenue would include training a larger classifier to recognise when there is no event present. Such a network could then feed back to the trigger for better detection accuracy, also enabling online learning. The proposed trigger establishes a novel, effective approach to reducing computational and power costs in audio processing while also demonstrating the potential to improve classification accuracy for SED. Its lightweight design makes it well suited for edge deployment, enabling local, real-time processing. 6. Funding This research was supported by the UK Government through the EPSRC Edgy Organism project (EP/Y030133/1). References L. F. Abbott (1999) Lapicque’s introduction of the integrate-and-fire model neuron (1907). Brain Research Bulletin 50 (5–6), p. 303–304. External Links: Document Cited by: §2.1. S. Adavanne, G. Parascandolo, P. Pertilä, T. Heittola, and T. Virtanen (2017) Sound event detection in multichannel audio using spatial and harmonic features. External Links: 1706.02293, Link Cited by: §1. C. M. Bishop (2006) Pattern recognition and machine learning. Springer. Cited by: §3.3. P. Blouw, X. Choo, E. Hunsberger, and C. Eliasmith (2019) Benchmarking keyword spotting efficiency on neuromorphic hardware. In Proceedings of the 7th Annual Neuro-Inspired Computational Elements Workshop, NICE ’19. External Links: Link, Document Cited by: §1. E. Cakir and T. Virtanen (2017) Convolutional recurrent neural networks for rare sound event detection. Technical report DCASE2017 Challenge. Cited by: §2.3, §2.3. G. Cerutti, R. Andri, L. Cavigelli, E. Farella, M. Magno, and L. Benini (2020) Sound event detection with binary neural networks on tightly power-constrained iot devices. In Proceedings of the ACM/IEEE International Symposium on Low Power Electronics and Design, ISLPED ’20, p. 19–24. External Links: Document Cited by: §1. B. Cramer, Y. Stradmann, J. Schemmel, and F. Zenke (2022) The heidelberg spiking data sets for the systematic evaluation of spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems 33 (7), p. 2744–2757. External Links: ISSN 2162-2388, Link, Document Cited by: §1. A. Dang, T. Vu, and J. Wang (2017) Deep learning for DCASE2017 challenge. Technical report DCASE2017 Challenge. Cited by: §2.3, §2.3, §3.3. D. De Benito-Gorrón, D. Ramos, and D. T. Toledano (2021) A multi-resolution crnn-based approach for semi-supervised sound event detection in dcase 2020 challenge. IEEE Access 9 (), p. 89029–89042. External Links: Document Cited by: §1. J. K. Eshraghian, M. Ward, E. Neftci, X. Wang, G. Lenz, G. Dwivedi, M. Bennamoun, D. S. Jeong, and W. D. Lu (2023) Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE 111 (9), p. 1016–1054. Cited by: §2.1. I. Goodfellow, Y. Bengio, and A. Courville (2016) Deep learning. MIT Press. Cited by: §3.3. S. Han, J. Pool, J. Tran, and W. J. Dally (2015) Learning both weights and connections for efficient neural networks. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 1, NIPS’15, Cambridge, MA, USA, p. 1135–1143. Cited by: §4.1.3, Table 5, Table 5. S. Heydari and Q. H. Mahmoud (2025) Tiny machine learning and on-device inference: a survey of applications, challenges, and future directions. 25 (10). External Links: Link, ISSN 1424-8220, Document Cited by: §3.2. M. Ji, Z. Wang, R. Yan, Q. Liu, S. Xu, and H. Tang (2023) SCTN: event-based object tracking with energy-efficient deep convolutional spiking neural networks. Volume 17. External Links: Link, Document, ISSN 1662-453X Cited by: §4.1.3. Y. Kawaguchi and T. Endo (2017) How can we detect anomalies from subsampled audio signals?. In 2017 IEEE 27th International Workshop on Machine Learning for Signal Processing (MLSP), Vol. , p. 1–6. External Links: Document Cited by: §3.1. S. Kim, S. Park, B. Na, and S. Yoon (2020) Spiking-yolo: spiking neural network for energy-efficient object detection. 34, p. 11270–11277. External Links: Document Cited by: §4.1.3. N. Koga, Y. Bando, and K. Imoto (2024) LEAD dataset: how can labels for sound event detection vary depending on annotators?. In 2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Vol. , p. 1–6. External Links: Document Cited by: §3.4.1, §3.4, §4.1.1. E. Lemaire, L. Cordone, A. Castagnetti, P. Novac, J. Courtois, and B. Miramond (2022) An analytical estimation of spiking neural networks energy efficiency. In Neural Information Processing, p. 574–587. External Links: Document Cited by: §4.1.3. Y. Li and X. Li (2017) The SEIE-SCUT systems for IEEE AASP challenge on DCASE 2017: deep learning techniques for audio representation and classification. Technical report DCASE2017 Challenge. Cited by: §2.3. H. Lim, J. Park, and Y. Han (2017) Rare sound event detection using 1D convolutional recurrent neural networks. Technical report DCASE2017 Challenge. Cited by: §2.3, §2.3. I. López-Espejo, Z. Tan, J. H. L. Hansen, and J. Jensen (2022) Deep spoken keyword spotting: an overview. 10 (), p. 4169–4199. External Links: Document Cited by: §3.1. R. Lu (2017) BIDIRECTIONAL gru for sound event detection. External Links: Link Cited by: §1. W. Maass (1997) Networks of spiking neurons: the third generation of neural network models. Neural Networks 10 (9), p. 1659–1671. External Links: ISSN 0893-6080, Document, Link Cited by: §1, §2.1. G. Matheron (1975) Random sets and integral geometry. Wiley. Cited by: §3.2. A. Mesaros, A. Diment, B. Elizalde, T. Heittola, E. Vincent, B. Raj, and T. Virtanen (2019) Sound event detection in the DCASE 2017 challenge. IEEE/ACM Transactions on Audio, Speech, and Language Processing. Note: In press External Links: Document, ISSN 2329-9290 Cited by: §2.3. A. Mesaros, T. Heittola, A. Diment, B. Elizalde, A. Shah, E. Vincent, B. Raj, and T. Virtanen (2017) DCASE 2017 Challenge setup: Tasks, datasets and baseline system. In DCASE 2017 - Workshop on Detection and Classification of Acoustic Scenes and Events, Munich, Germany. External Links: Link Cited by: §1, §2.2, §2.3, §3.1. A. Mesaros, T. Heittola, and T. Virtanen (2016) Metrics for polyphonic sound event detection. Applied Sciences 6 (6). External Links: Link, ISSN 2076-3417, Document Cited by: §2.2. E. O. Neftci, H. Mostafa, and F. Zenke (2019) Surrogate gradient learning in spiking neural networks: bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine 36 (6), p. 51–63. External Links: Document Cited by: §2.1. H. Phan, M. Krawczyk-Becker, T. Gerkmann, and A. Mertins (2017) DNN and CNN with weighted and multi-task loss functions for audio event detection. Technical report DCASE2017 Challenge. Cited by: §2.3. A. Ravichandran and S. Das (2017) Bosch rare sound events detection systems for DCASE2017 challenge. Technical report DCASE2017 Challenge. Cited by: §2.3. J. Salamon, C. Jacoby, and J. P. Bello (2014) A dataset and taxonomy for urban sound research. In Proceedings of the 22nd ACM International Conference on Multimedia, M ’14, New York, NY, USA, p. 1041–1044. External Links: ISBN 9781450330633, Link, Document Cited by: §3.1. J. Salamon, D. MacConnell, M. Cartwright, P. Li, and J. P. Bello (2017) Scaper: a library for soundscape synthesis and augmentation. In 2017 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), Vol. , p. 344–348. External Links: Document Cited by: §3.1. F. Schmid, P. Primus, T. Morocutti, J. Greif, and G. Widmer (2024) IMPROVING audio spectrogram transformers for sound event detection through multi-stage training. Technical report DCASE2024 Challenge. Cited by: §1. J. Serra (1982) Image analysis and mathematical morphology. Academic Press, London. Cited by: §3.2. J. Sevilla, L. Heim, A. Ho, T. Besiroglu, M. Hobbhahn, and P. Villalobos (2022) Compute trends across three eras of machine learning. In 2022 International Joint Conference on Neural Networks (IJCNN), Vol. , p. 1–8. External Links: Document Cited by: §1. D. Stowell, D. Giannoulis, E. Benetos, M. Lagrange, and M. D. Plumbley (2015) Detection and classification of acoustic scenes and events. IEEE Transactions on Multimedia 17 (10), p. 1733–1746. External Links: Document Cited by: §1. G. Tang, K. Vadivel, Y. Xu, R. Bilgic, K. Shidqi, P. Detterer, S. Traferro, M. Konijnenburg, M. Sifalakis, G. van Schaik, and A. Yousefzadeh (2023) SENECA: building a fully digital neuromorphic processor, design trade-offs and challenges. Volume 17. External Links: Link, Document, ISSN 1662-453X Cited by: §4.1.3, Table 5, Table 5. M. C. W. Van Rossum (2001) A novel spike distance. Neural Comput. 13 (4), p. 751–763. External Links: ISSN 0899-7667 Cited by: §3.2. S. Venkatesh, D. Moffat, and E. R. Miranda (2022) You only hear once: a yolo-like algorithm for audio segmentation and sound event detection. Applied SciencesIEEE AccessSensorsApplied SciencesProceedings of the AAAI Conference on Artificial IntelligenceFrontiers in NeuroscienceFrontiers in Neuroscience 12 (7). External Links: Link, ISSN 2076-3417 Cited by: §3.4, §4.1.1. J. Wang, W. Zhang, and J. Liu (2017) Transfer learning based DNN-HMM hybrid system for rare sound event detection. Technical report DCASE2017 Challenge. Cited by: §2.3. L. Wang (2025) PRE-trained model enhanced anomalous sound detection system for dcase2025 task2. Technical report DCASE2025 Challenge. Cited by: §1. Y. Wang, Y. Zheng, Y. Zhang, Y. Xie, S. Xu, Y. Hu, and L. He (2021) Unsupervised anomalous sound detection for machine condition monitoring using classification-based methods. 11 (23). External Links: Link, ISSN 2076-3417 Cited by: §1. Q. Wu, Q. Zhang, C. Tan, Y. Zhou, and C. Sun (2024) Point-to-spike residual learning for energy-efficient 3d point cloud classification. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, p. 6092–6099. External Links: Document Cited by: §4.1.3. xiaoju ye (2023) External Links: Link Cited by: Figure 2. J. Yang (2025) A two stage fusion anomaly detection approach for task2. Technical report DCASE2025 Challenge. Cited by: §1. Y. Yang, W. Zhang, and P. Li (2021) Backpropagated neighborhood aggregation for accurate training of spiking neural networks. In International Conference on Machine Learning, External Links: Link Cited by: §3.2. Z. Ye, X. Wang, H. Liu, Y. Qian, R. Tao, L. Yan, and K. Ouchi (2021) Sound event detection transformer: an event-based end-to-end model for sound event detection. External Links: 2110.02011, Link Cited by: §3.4, §4.1.1. W. Zhang and P. Li (2020) Temporal spike sequence learning via backpropagation for deep spiking neural networks. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY, USA. External Links: ISBN 9781713829546 Cited by: §3.2.