Spiking Heidelberg Digits (SHD)
- Spiking Heidelberg Digits (SHD) is a spike-based speech benchmark defined by biologically inspired auditory processing and precise temporal dynamics for digit classification.
- The dataset standardizes training and evaluation protocols by converting spoken digit audio into spike trains, enhancing reproducibility across SNN studies.
- SHD serves as a dual-purpose tool by supporting both algorithmic advancements in temporal coding and hardware performance testing on neuromorphic platforms.
Spiking Heidelberg Digits (SHD) is a spike-based speech benchmark introduced within the Heidelberg spiking datasets as a free, high-fidelity, word-level aligned dataset for digit classification, designed to support systematic evaluation of spiking neural networks (SNNs) and neuromorphic hardware (Cramer et al., 2019). It consists of spoken digits recorded in English and German and converted into spike trains by a biologically inspired auditory pipeline, and it has subsequently become a standard testbed for temporal coding, surrogate-gradient training, delay-based computation, and hardware deployment studies (Cramer et al., 2019).
1. Origin and dataset construction
SHD was created to address the lack of a widely accepted spike-based benchmark that is free, publicly available, non-saturated, and useful for both software and hardware evaluation (Cramer et al., 2019). The original dataset paper positions SHD against benchmarks that either require user-defined spike conversion, are nearing saturation, or are poorly suited to temporal coding studies, and it frames SHD as a reference point for comparing SNN training algorithms and neuromorphic implementations (Cramer et al., 2019).
The source material is spoken-digit audio. The dataset contains digits $0$–$9$, recorded in English and German, and the recordings were produced in a sound-shielded room using three microphones, digitized at $48$ kHz with $24$-bit precision, then manually pre-selected, cut, and mastered (Cramer et al., 2019). The original paper reports approximately $10$k recordings, more precisely total digit recordings, from $12$ speakers, with $6$ female and $6$ male speakers, ages $21$–$9$0, mean age $9$1 (Cramer et al., 2019).
The audio-to-spike conversion is explicitly neurophysiology-inspired. The pipeline comprises a basilar membrane model, a hair cell model, and a bushy cell layer (Cramer et al., 2019). At a high level, the basilar membrane model decomposes audio into frequency components, the hair cell model converts those components into firing rates with refractory effects, and the bushy cell layer combines multiple hair-cell spike trains to increase phase locking (Cramer et al., 2019). The purpose of this construction is to reduce user-specific preprocessing choices and make downstream comparisons more reproducible (Cramer et al., 2019).
2. Representation, splits, and task conventions
In the original description, SHD is an event-based dataset stored as spike trains in HDF5 format, with audio as input modality and digit classification as the target task (Cramer et al., 2019). The original paper describes the benchmark as a $9$2-way digit classification problem over digits $9$3–$9$4 in English and German, and it states that the train/test split was designed to test cross-speaker generalization by holding out $9$5 speakers exclusively for test and filling the remaining test set with $9$6 of trials from speakers also present in training (Cramer et al., 2019).
Later work often instantiates SHD with a canonical split of $9$7 training samples and $9$8 test samples, $9$9 input channels, and $48$0 output classes, typically corresponding to digits $48$1–$48$2 spoken in both German and English (Takaghaj et al., 2024). Spiker+ uses SHD in exactly that form, with $48$3 inputs, $48$4 spike time-steps, custom encoding from the SHD benchmark reference, and a fully connected recurrent spiking neural network of architecture $48$5-$48$6-$48$7 (Carpegna et al., 2024). Sparse Spiking Gradient Descent instead uses $48$8 input neurons, $48$9 output classes, two hidden layers of $24$0 spiking neurons each, and a simulation horizon of $24$1 time steps with $24$2 ms (Perez-Nieves et al., 2021). Delay-oriented hardware work often bins the $24$3 inputs to $24$4 channels, for example by binning every $24$5 input neurons and every $24$6 ms before training a $24$7 network (Chen et al., 3 Nov 2025). The hls4ml FPGA workflow applies a different fixed-window preprocessing: binning spikes with $24$8 ms bins over $24$9 time steps and reducing $10$0 channels to $10$1 features by averaging consecutive groups of $10$2 channels (Dillon, 8 Jun 2026).
These conventions are not identical. A plausible implication is that SHD functions in practice as a stable underlying spike-based corpus with multiple downstream protocol choices rather than as a single immutable preprocessing standard.
3. Temporal structure and the rate-versus-timing question
A central property of SHD is that it is not a static-image benchmark. Later work repeatedly treats it as a temporal, event-driven auditory task in which the time dimension is essential to classification (Carpegna et al., 2024). The original dataset paper already showed that temporal information is critical: when temporal structure is removed and classifiers are trained only on spike-count vectors, a linear SVM reaches only slightly above $10$3 on SHD and the best nonlinear SVMs reach about $10$4, whereas temporal classifiers perform much better, with an LSTM reaching $10$5 (Cramer et al., 2019). This result established SHD as a benchmark where leveraging spike timing information is essential for good classification accuracy (Cramer et al., 2019).
At the same time, later analysis shows that raw SHD contains substantial spike-count biases across channels and classes, so vanilla SHD is not a pure test of temporal coding (Yu et al., 21 Jul 2025). To isolate timing, that work constructs timing-normalized variants, denoted whole $10$6 part $10$7 norm, and reports that the class-balanced chance level for SHD is $10$8 (Yu et al., 21 Jul 2025). On the timing-only SHD-norm benchmark at $10$9, a surrogate-gradient SNN reaches 0 and an SNN with learnable delays reaches 1, both above chance, while a rate-only MLP baseline collapses to chance (Yu et al., 21 Jul 2025). The same study also reports strong sensitivity to temporal perturbations such as per-spike jitter, per-neuron jitter, spike deletion, and time reversal, supporting the interpretation that SHD contains both rate-confounded information and precise temporal structure (Yu et al., 21 Jul 2025).
This combination has made SHD important in two distinct senses. First, it is a practical benchmark for end-to-end speech classification with spikes. Second, it is a methodological instrument for separating rate-based from timing-based computation.
4. Role in SNN algorithm development
SHD is widely used as a stress test for training methods because it is long, temporally structured, and less trivially sparse than common neuromorphic vision datasets (Perez-Nieves et al., 2021). It therefore appears in work on sparse backpropagation, trainable thresholds, exact-gradient methods, explicit delay mechanisms, skip connections, and multi-scale temporal modeling.
Representative reported SHD results include the following.
| Method | Setup | Reported SHD result |
|---|---|---|
| Sparse Spiking Gradient Descent (Perez-Nieves et al., 2021) | 2-3-4-5 fully connected SNN | 6 test accuracy; up to 7 backward speedup in the default setup; nearly 8 speedup and about 9 memory improvement at $12$0 |
| Rouser (Takaghaj et al., 2024) | $12$1-$12$2-$12$3-$12$4 feedforward SNN with trainable thresholds | $12$5 vs $12$6 for Sparse BP |
| Eventprop with loss shaping (Nowotny et al., 2022) | recurrent SHD network with augmentations and delay line | $12$7 test accuracy; $12$8 maximum cross-validation accuracy |
| TSkips (Malettira et al., 2024) | deeper MLP backbone with explicit temporal delay skip connections | $12$9 for Baseline-2 with forward + backward TSkips |
| State-space delay SNN (Karilanova et al., 1 Dec 2025) | adLIF, $6$0, Uniform non-trainable delay matrix | $6$1 with $6$2M parameters |
| TR/NAR over SNN-Delays (Zhang et al., 2024) | SNN-Delays + Non-Aligned Residual + TR-no | $6$3 |
These values are not directly comparable in a strict benchmark sense because the papers use different preprocessing pipelines, architectures, validation procedures, and, in some cases, distinct problem formulations. That caveat is explicit in the literature itself. Eventprop, for example, emphasizes a strict $6$4-fold leave-one-speaker-out cross-validation protocol and notes that some higher headline SHD figures elsewhere may depend on different validation methodology (Nowotny et al., 2022).
Several methodological themes recur across these results. Sparse Spiking Gradient Descent argues that most gradient computations are unnecessary because most neurons are far from threshold most of the time, even on SHD, where average activity is only about $6$5 in the first hidden layer and $6$6 in the second hidden layer (Perez-Nieves et al., 2021). Rouser promotes firing threshold from fixed hyperparameter to trainable parameter and reports better convergence, fewer dead neurons, and reduced threshold grid search (Takaghaj et al., 2024). Eventprop shows that exact gradients alone are insufficient unless the loss is shaped appropriately for sequential speech tasks (Nowotny et al., 2022). Delay-based methods, whether implemented as explicit axonal delays, state-space memory variables, or delayed skip connections, consistently treat SHD as evidence that temporal dependencies can be modeled without relying only on conventional recurrence (Malettira et al., 2024).
A related sparsity-oriented line uses Linearized Bregman Iterations and AdaBreg on a binned SHD configuration with a recurrent middle layer and learned axonal delays, reporting $6$7 test accuracy with learning-rate scheduling, $6$8 without scheduling, and about $6$9 reduction in active parameters relative to Adam-trained baselines (Windhager et al., 17 Mar 2026).
5. Hardware benchmarking and deployment
SHD is also a hardware benchmark. Because the input is already spike-based and temporally structured, it is frequently used to assess whether FPGA or SoC implementations can preserve temporal processing while meeting latency, power, and resource constraints.
Representative deployment results are summarized below.
| System | Platform and model | Reported SHD deployment |
|---|---|---|
| Spiker+ (Carpegna et al., 2024) | XA7Z020, FCR SNN, $6$0-$6$1-$6$2, II-order LIF | $6$3, $6$4 ms latency, $6$5 W, $6$6 logic cells, $6$7 BRAM |
| Synaptic-delay SNN processor (Chen et al., 3 Nov 2025) | PYNQ Z2 SoC, $6$8, feedforward LIF + delays | $6$9, $21$0 ms/sample, $21$1 samples/s, $21$2 mW |
| Event-graph SoC FPGA (Nakano et al., 9 Mar 2025) | ZCU104, quantized PointNetConv-style event graph | $21$3, $21$4s after the last event, $21$5 W PL power |
| hls4ml SNN inference (Dillon, 8 Jun 2026) | xczu7ev-ffvc1156-2-e, $21$6-$21$7-$21$8 fixed-window LIF SNN | $21$9 HLS test accuracy at ap_fixed\<10,4>; $9$00s full-window latency |
These results emphasize different objectives. Spiker+ explicitly presents SHD as a harder benchmark than MNIST because SHD requires a more complex neuron model, a recurrent architecture, and careful handling of temporal spike dynamics (Carpegna et al., 2024). The synaptic-delay processor instead treats SHD as a keyword-spotting task and uses a spiking ring buffer to emulate delays efficiently in hardware, reporting $9$01 peak pre-quantization accuracy and $9$02 after $9$03-bit symmetric quantization (Chen et al., 3 Nov 2025). The event-graph FPGA work uses SHD not to validate an SNN but to argue that sparse event-graph neural networks can be a hardware-friendly alternative for cochlea-derived audio, achieving $9$04 floating-point and $9$05 quantized accuracy on SHD (Nakano et al., 9 Mar 2025). The hls4ml extension demonstrates that a PyTorch-trained SNN can be converted into fixed-point FPGA firmware with close agreement between quantized software and HLS simulation, with reported agreement up to $9$06 depending on precision (Dillon, 8 Jun 2026).
A recurring pattern across these studies is that SHD is used not merely as a classification benchmark but as a joint algorithm-hardware probe of temporal processing, statefulness, quantization tolerance, and memory organization.
6. Variants, extended uses, and recurring issues
SHD has been repurposed well beyond ordinary supervised classification. In multimodal fusion, it serves as the auditory branch paired with N-MNIST; one such study uses only the ten English digit classes from SHD and reports $9$07 for the auditory-only branch and $9$08 for late fusion with the visual branch (Bjorndahl et al., 2024). In continual learning, Replay4NCL uses SHD in a class-incremental setting with $9$09 pre-training classes and one continual-learning class, reporting old-task Top-1 accuracy of $9$10 versus $9$11 for a SpikingLR baseline, along with $9$12 latency speed-up, $9$13 latent memory saving, and $9$14 energy saving (Minhas et al., 21 Mar 2025). In federated learning, a subset of SHD restricted to digits $9$15–$9$16 yields a $9$17-class task with $9$18 training and $9$19 testing samples; the reported conclusion is that satisfactory performance depends on a trade-off between random masking and client drop probability (Chaki et al., 2023). In integer-only online training, SHD is used to test whether low-precision temporal learning remains viable; a recurrent SNN reaches $9$20 in mixed precision, close to a $9$21 BPTT RSNN baseline (Gomez et al., 8 Sep 2025).
SHD also appears in nonstandard recognition settings. A bounded-delay motif-detection study states that motifs of arbitrary length extracted from SHD can be recognized by chaining sub-motif detectors, with about $9$22 correct detection rate in the presence of ten simultaneous motifs from SHD and up to $9$23 for five motifs (Kronland-Martinet et al., 19 Nov 2025). Because the provided document for that work is a cover letter rather than the full paper, it omits SHD preprocessing, architecture details, and formulae; this limits the technical specificity of that particular use case (Kronland-Martinet et al., 19 Nov 2025).
Several recurring issues accompany SHD research. One is class convention: the original dataset paper describes a $9$24-way digit task over digits $9$25–$9$26 in English and German, whereas many later works instantiate $9$27 output classes corresponding to language-specific digit labels (Cramer et al., 2019). Another is rate confounding: raw SHD permits nontrivial performance from spike counts alone, so derived variants such as SHD-norm have been proposed when the research question specifically targets precise timing (Yu et al., 21 Jul 2025). A third is preprocessing heterogeneity: channel binning, temporal framing, output conventions, and readout rules vary substantially across papers. This suggests that SHD is best understood as a benchmark family anchored by a common cochlea-derived spoken-digit corpus, with downstream protocols chosen to probe different aspects of temporal computation.
