Fig. 3: Audio localization with STHT, RZCC encoding and DoA inference with a Spiking Neural Network (SNN). | Communications Engineering

Fig. 3: Audio localization with STHT, RZCC encoding and DoA inference with a Spiking Neural Network (SNN).

From: Low-power Spiking Neural Network audio source localisation using a Hilbert Transform audio event encoding scheme

Fig. 3

a The pipeline for SNN implementation of our Hilbert beamforming and DoA estimation, combining Short-Time Hilbert transform; Zero-crossing conjugate encoding; analytically-derived beamforming weights; and Leaky-Integrate-and-Fire (LIF) spiking neurons for power accumulation and DoA estimation. b, c Beam patterns for SNN STHT RZCC beamforming, for narrowband (b; F = 2 kHz) and wideband (c; FC = 2 kHz) signals. d, e Beam power and DoA estimates for noisy narrowband signals (d) and for noisy encoded speech (e). Dashed lines: estimated DoA. f, g DoA estimation error for noisy narrowband signals (f) and noisy speech (g). Dashed line: 1°. Annotations: Mean Absolute Error (MAE). Box plot: centre line: median; box limits: quartiles; whiskers: 1.5 × inter-quartile range; points: outliers. n = 100 random trials.

Back to article page