Two techniques that improve spiking neural networks for speech classification. Temporal Reconstruction rebuilds the temporal dimension of an audio spectrogram at multiple resolutions, mimicking the brain's hierarchical multi-time-scale processing so the network captures information at different time scales. Non-Aligned Residual enables residual (skip) connections between audio representations of different time lengths, which single-resolution SNN pipelines otherwise cannot use. Together they set state-of-the-art accuracy on the Spiking Speech Commands and Spiking Heidelberg Digits benchmarks while reducing processing time and energy.
Spiking neural networks (SNNs) encode information in the timing and frequency of discrete spikes, making them energy-efficient and naturally suited to temporal data such as speech, but most SNN speec…