Spiking Speech Commands Benchmark

• Spiking Speech Commands (SSC) is a neuromorphic benchmark that converts audio recordings into spike trains using an artificial cochlear model for 35-class spoken-word recognition.
• SSC employs standardized preprocessing by reducing 700 input neurons to 140 through binning and discretizing 1-second samples into 100 timesteps for consistent evaluations.
• Various model families, from delay-learning SNNs to transformer-based architectures, leverage SSC to probe temporal coding, energy efficiency, and performance trade-offs.

Spiking Speech Commands (SSC) is a neuromorphic speech-command benchmark derived from Google Speech Commands by converting audio recordings into spike trains with an artificial cochlear model, or, in one description, “an artificial model of the inner ear and parts of the ascending auditory pathway.” In the literature surveyed here, SSC is used as a 35-way spoken-command classification task over event-based auditory inputs, and later papers report a total of 105,829 samples with train/validation/test partitions of 75,466/9,981/20,382 (Deckers et al., 2023, Zhang et al., 2024, Wang et al., 11 Nov 2025). SSC has become a central testbed for spiking neural networks because it combines large-vocabulary command recognition with spike-native temporal structure, and it is commonly studied alongside the smaller Spiking Heidelberg Digits (SHD) benchmark and the non-spiking Google Speech Commands (GSC) dataset (Malettira et al., 2024).

1. Benchmark identity and scope

SSC is consistently described as the spiking counterpart of Google Speech Commands: both target spoken-word classification, but SSC presents the input as spikes rather than waveforms or spectrograms (Wang et al., 11 Nov 2025). One paper characterizes SSC as the larger companion to SHD, with 100,000 recordings of spoken words from various speakers and a top-1 classification task over all 35 classes, while later transformer-based studies give the exact count 105,829 and the split 75,466 / 9,981 / 20,382 (Malettira et al., 2024, Wang et al., 11 Nov 2025).

The task is harder than SHD because SSC inherits the larger-vocabulary command setting of GSC rather than a digit-only vocabulary. Several papers make this comparison explicit: SHD is a spoken-digit task, whereas SSC and GSC both have 35 classes and therefore represent a more challenging speech recognition problem (Deckers et al., 2023). In practice, SSC is used both for absolute benchmark comparison and for probing questions about temporal coding, delay learning, sparsity, and energy-aware neuromorphic inference (Hammouamri et al., 2023, Windhager et al., 17 Mar 2026).

A recurring representation-level description is the event form $(x,t)$, where $x$ is the spike unit or channel and $t$ is the timestamp (Malettira et al., 2024). This spike-native formulation is what distinguishes SSC from GSC-based spiking pipelines that begin from conventional audio features and only later convert them into spikes.

2. Representation, preprocessing, and temporal discretization

A strong convention across SSC papers is that the original spike representation has 700 input neurons or input channels, and many models reduce this to 140 by binning every five neurons (Deckers et al., 2023, Hammouamri et al., 2023). This 700-to-140 reduction appears in recurrent, feed-forward, delay-learning, sparse-training, and transformer-based systems alike, which makes it one of the closest things SSC has to a de facto preprocessing standard (Windhager et al., 17 Mar 2026, Wang et al., 11 Nov 2025).

Another common convention is duration normalization to roughly 1 second. One paper states that all samples are zero-padded and aligned to 1 second, then binned into 100 timesteps using 10 ms bins (Deckers et al., 2023). DCLS-Delays likewise uses $\Delta t = 10$ ms with zero right-padding to equalize durations in a batch (Hammouamri et al., 2023). Later work keeps the same 1000 ms duration but varies the temporal resolution: SpikCommander uses $T = 1000 / \Delta t$ with $\Delta t \in \{1,4,10\}$ ms and emphasizes $T=100$ for fair comparison, while TSkips reports sequence length $T=99$ for SHD/SSC experiments (Wang et al., 11 Nov 2025, Malettira et al., 2024).

These discretization choices are not merely bookkeeping. They determine the number of SNN simulation steps and therefore directly affect latency, operation count, and the feasibility of modeling long-range temporal structure. This suggests that “SSC performance” is partly a property of the benchmark and partly a property of how the spike stream is temporally binned and padded.

Spike-domain augmentation is also present in later SSC systems. Both SpikeSCR and SpikCommander use event-drop-style augmentation with drop-by-time and drop-by-neuron, with SSC parameters reported as drop proportion 50%, time drop size 10%, and neuron drop size 10 (Wang et al., 2024, Wang et al., 11 Nov 2025). Earlier work instead explored augmentations such as shift, blend, random dilation, and ID jitter; in the Eventprop study, the strongest SSC combinations were delay+shift and delay+shift+blend rather than blend alone (Nowotny et al., 2022).

3. Model families and learning paradigms

SSC has supported several distinct architectural lines. One influential family treats delays as first-class trainable parameters. DCLS-Delays models each delayed synapse as a sparse 1D temporal convolution and learns delay positions with Dilated Convolutions with Learnable Spacings; on SSC it uses 2 or 3 hidden fully connected layers, hidden size 512, vanilla LIF neurons, dropout 0.25, and a maximum delay of 300 ms (Hammouamri et al., 2023). A closely related line co-learns weights, synaptic delays, and neuronal adaptation in a 2 hidden-layer fully connected feed-forward SNN with 512 AdLIF+ neurons per hidden layer (Deckers et al., 2023).

A second family emphasizes recurrent delayed-LIF architectures. The sparse-optimization paper based on DelRec uses an SSC network of the form $$140 \rightarrow 256 \rightarrow 256 \rightarrow 256 \rightarrow 35$$, with three recurrent hidden layers, LIF neurons, and learned axonal delays in the recurrent path; sparsity is then imposed through Linearized Bregman Iterations, implemented as AdaBreg (Windhager et al., 17 Mar 2026). Exact-gradient training via Eventprop uses a different recurrent SNN regime: a single recurrent hidden layer, delay-line input expansion, voltage-based readout, and a loss-shaped objective based on the exponentially weighted integral of output voltages (Nowotny et al., 2022).

A third family replaces recurrence-heavy designs with architectural mechanisms for long-range temporal context. TSkips augments fully connected spiking MLPs with explicit delayed forward and backward skip connections found by training-free NAS; its best SSC model is a searched 8-layer spiking MLP with both forward and backward delayed skips (Malettira et al., 2024). More recent work moves further toward spike-driven sequence modeling with hybrid or transformer-like designs. SpikeSCR combines spiking self-attention with a local separable gated-convolution path in a global-local hybrid architecture (Wang et al., 2024). SpikCommander uses a fully spike-driven transformer with multi-view spiking temporal-aware self-attention (MSTASA) and a spiking contextual refinement channel MLP (SCR-MLP), explicitly targeting complementary local and global temporal dependencies in SSC (Wang et al., 11 Nov 2025).

Across these families, a consistent pattern is that SSC rewards mechanisms for explicit temporal alignment, delayed interaction, or multiscale temporal aggregation. This suggests that SSC is not just a larger SHD; it is a benchmark on which temporal inductive bias is central rather than incidental.

4. Temporal coding, normalization, and SSC-derived variants

A major conceptual development around SSC is the distinction between standard SSC and SSC-derived variants that remove rate information. One paper argues that raw SSC contains both temporal structure and substantial spike-count bias, so strong performance on standard SSC does not by itself prove temporal coding (Yu et al., 21 Jul 2025). To isolate timing, that work constructs three variants: whole SSC, part SSC, and norm SSC.

The construction starts from a binary tensor $X \in \{0,1\}^{M \times N \times T}$. In the part stage, neurons and some samples are filtered so that retained neurons have at least a minimum spike count in all remaining samples, using threshold $x$0 and $x$1 (Yu et al., 21 Jul 2025). Classes are then balanced by random downsampling. In the norm stage, each retained neuron is forced to have the same number of spikes in every sample by subsampling to the per-neuron minimum count across filtered samples. The result is a timing-only benchmark in which all per-neuron spike-count information across samples is removed (Yu et al., 21 Jul 2025).

This variant construction yields one of the clearest statements about what SSC contains. On SSC-norm, a rate-only MLP falls to chance = 2.8%, while a surrogate-gradient SNN without delays reaches 14.69% and the same SNN with learnable delays reaches 32.60% (Yu et al., 21 Jul 2025). The paper interprets this as direct evidence that SSC contains discriminative temporal information beyond firing rates and that delay learning materially improves access to that information. It also reports that with increasing timing corruption $x$2, both SNNs degrade and the delay-equipped model shows a sharper drop, which the authors read as stronger reliance on precise spike timing rather than a mere count-based shortcut (Yu et al., 21 Jul 2025).

A plausible implication is that standard SSC mixes at least two computational problems: classification from rate/statistics and classification from timing. The whole/part/norm progression was introduced precisely to separate those two sources of performance.

5. Accuracy, sparsity, and efficiency trade-offs

Direct comparison across SSC papers requires caution because training budgets, time steps, architectures, augmentation, and even benchmark variants differ. Even so, the literature shows a clear progression from mid-70% recurrent SNNs to low-80% delayed, hybrid, and transformer-like systems under standard 35-class SSC settings.

Model family	Reported SSC setting	Accuracy
AdLIF+ with synaptic delays	2 hidden layers, 512 units	79.81%
DCLS-Delays (3L-2KC)	100 steps	80.69%
SNN-Delays + TR/NAR	3L-2KC backbone	81.02%
AdaBreg on DelRec	sparse recurrent delayed-LIF SNN	81.86%
SpikCommander (2L-16-256)	100 steps	83.49%
SpikeSCR (2L-16-256) + KDCL	100 steps	83.69%

These values come from distinct model families and should be read as representative rather than perfectly like-for-like (Deckers et al., 2023, Hammouamri et al., 2023, Zhang et al., 2024, Windhager et al., 17 Mar 2026, Wang et al., 11 Nov 2025, Wang et al., 2024).

Several papers report that SSC continues to improve when the temporal budget is increased. SpikeSCR rises from 82.79% at 100 steps to 84.47% at 200 steps and 85.57% at 500 steps, whereas DCLS-Delays degrades from 81.08% at 100 to 78.76% at 500 in the same comparison (Wang et al., 2024). SpikCommander also benefits from longer sequences, increasing from 83.49% at 100 steps to 85.52% at 200 and 85.98% at 250 (Wang et al., 11 Nov 2025). This suggests that recent attention-based SSC models scale more favorably with time-step count than earlier delay-convolution baselines.

Efficiency claims are more heterogeneous. TSkips reaches 80.23% on SSC with 1.42M parameters, compared with 80.69% for DCLS-Delays at 2.5M parameters, so it is slightly less accurate but smaller (Malettira et al., 2024). AdaBreg preserves SSC accuracy close to an Adam-trained DelRec baseline while reducing the number of active parameters by about 50%; the best reported SSC numbers are 82.58% for the Adam baseline and 81.86% for AdaBreg (Windhager et al., 17 Mar 2026). SpikeSCR’s curriculum-learning knowledge distillation compresses the time budget from 100 to 40 steps and reduces SSC energy from 0.0314 mJ to 0.0142 mJ, a 54.8% reduction, while maintaining 80.25% accuracy, comparable to DCLS-Delays (2L-2KC) at 80.16% with 100 steps (Wang et al., 2024). SpikCommander reports 0.0173G SOPs and 0.0180 mJ on SSC for its 2-layer model, corresponding to 50.3% fewer SOPs, 42.7% lower energy, and 35.5% fewer parameters than the compared SpikeSCR baseline (Wang et al., 11 Nov 2025).

Not all training regimes are equally competitive. Eventprop, despite exact gradients and substantial engineering, reports a best SSC validation accuracy of 74.7% and a corresponding test accuracy of 74.1 ± 0.9%, which the authors characterize as good rather than state of the art (Nowotny et al., 2022). This places SSC as a benchmark on which optimizer choice, loss shaping, and augmentation strategy can be decisive.

6. Acronym ambiguity and adjacent command-recognition benchmarks

“SSC” is not unambiguous across the literature. In the paper titled “Speech Command Recognition in Computationally Constrained Environments with a Quadratic Self-organized Operational Layer”, SSC means Synthetic Speech Commands, specifically the Kaggle “synthetic-speech-commands-dataset”, described as an automatically generated text-to-speech counterpart of GSC (Soltanian et al., 2020). That work uses MFCCs and a conventional differentiable network, not spiking data, and it explicitly should not be read as a paper on the neuromorphic SSC benchmark (Soltanian et al., 2020).

There is also a broader ring of adjacent research on speech-command recognition with SNNs that does not evaluate on SSC directly. A supervised dilated convolutional LIF SNN on Google Speech Commands v1 reports 94.5% test accuracy with 0.13M parameters and low spiking activity, but it uses log-Mel features and the GSC task rather than SSC (Pellegrini et al., 2020). A learnable residual speech-to-spike encoder trained on GSC-v2 reaches 94.97% and frames its contribution as adaptive speech-to-spike encoding rather than SSC benchmarking (Anon et al., 17 Jun 2026). A neuromorphic keyword-spotting system using PDM-converted GSC reports 91.54% and compares this number to SSC literature without evaluating on SSC itself (Yarga et al., 2024). Early-decision spiking keyword spotting on GSC v2 and SC-100 studies latency-aware inference rather than the standard SSC benchmark, although its frame-synchronous design is highly relevant to practical SSC-style deployment (Song et al., 2024).

For this reason, careful acronym disambiguation is essential. In neuromorphic speech papers, SSC usually denotes Spiking Speech Commands; in some conventional audio-classification papers, it instead denotes Synthetic Speech Commands. The two should not be conflated, and benchmark comparisons are only meaningful when representation, split, and task protocol are aligned.