Adaptive speech-to-spike encoder boosts SNN accuracy to 94.97% on GSC-v2

The fundamental mismatch between continuous acoustic signals and discrete spike-based processing has long hindered neuromorphic speech systems. A new paper from Rahman Anon and Emon (accepted at Interspeech 2026) tackles this by introducing a learnable residual speech-to-spike encoder, trained end-to-end with a Recurrent Leaky Integrate-and-Fire (R-LIF) backbone. On the 35-class Google Speech Commands v2 benchmark, the system achieves up to 94.97% accuracy, significantly outperforming fixed encoder baselines. Remarkably, a parameter-efficient variant with only 35k parameters reaches 89.8%—matching or exceeding prior models that require an order of magnitude more parameters, demonstrating that learned encoders can drastically reduce computational overhead.

The authors probe the encoder's internal representations using linear probing and gradient-residual inspection, revealing that it does not aim for faithful signal reconstruction but instead learns task-aligned spike representations that maximize class separability. Additionally, the paper benchmarks bio-inspired credit assignment by comparing Direct Feedback Alignment (DFA) with standard surrogate-gradient backpropagation-through-time (BPTT) under identical architectures. DFA reaches 91.5% accuracy, only 3.5% below BPTT, while offering more hardware-friendly, local learning rules. This quantifies the performance trade-off for deploying neuromorphic audio in low-power, edge scenarios where bio-plausibility and energy efficiency are critical.