Adaptive Speech-to-Spike Encoding for Spiking Neural Networks

Researchers have developed a learnable speech-to-spike encoder for Spiking Neural Networks that reaches 94.97% accuracy on the Google Speech Commands v2 benchmark, according to a paper published on arXiv [1]. The encoder is jointly trained with a recurrent spiking backbone and achieves competitive performance even at a compact 35,000-parameter scale [1]. The work, authored by Taharim Rahman Anon and Jakaria Islam Emon, addresses what the paper describes as a “fundamental bottleneck for neuromorphic speech processing” — the mismatch between continuous acoustic signals and the discrete, event-driven nature of spiking neural networks [2][4]. Conventional systems rely on fixed spike encoders, forcing the downstream network to adapt to representations it cannot shape [1]. The proposed architecture replaces those fixed thresholds with learnable parameters, allowing the encoder and a Recurrent Leaky Integrate-and-Fire (R-LIF) backbone to be optimized end-to-end [4]. On the Google Speech Commands v2 dataset, a standard keyword-spotting benchmark, the full model achieves up to 94.97% accuracy [1][2]. A compact variant with only 35,000 parameters reaches 89.8%, which the authors note matches or exceeds prior baselines that require an order of magnitude more parameters [1][4]. The encoder does not attempt to faithfully reconstruct the input log-mel spectrogram. Instead, linear probing and gradient-residual analysis indicate it learns task-aligned spike representations that increase the linear separability of classes compared to fixed baselines [1][4]. The average encoder spike rate is reported at 6.56% [4]. The paper also benchmarks a bio-inspired credit-assignment method, Direct Feedback Alignment (DFA), against surrogate-gradient backpropagation through time (BPTT). Under identical architectures and training conditions, DFA reaches 91.5% accuracy, quantifying the performance trade-off of a learning rule considered more hardware-friendly for neuromorphic systems [1][2]. The approach builds on a broader effort to equip spiking neural networks with effective auditory front-ends. A 2023 study introduced Spiking-LEAF, a learnable auditory front-end that combined a trainable filter bank with a two-compartment spiking neuron model inspired by inner hair cells, and incorporated lateral inhibition and spike-rate regularization to improve encoding efficiency on keyword spotting and speaker identification tasks [5]. Recurrent neural networks, which underpin the R-LIF backbone, are designed for sequential data such as speech and maintain a hidden state that captures temporal dependencies across time steps [7]. The new encoder’s parameter efficiency and compatibility with bio-inspired learning rules may ease deployment on low-power neuromorphic hardware, where spike-based computation offers energy advantages over conventional deep learning accelerators [1][4].