From Speckle to Spectrum: Learned Inversion of a Passive Neuromorphic Acoustic Sensor

Sound doesn’t have to be measured electrically. We show that a neuromorphic event camera, watching the laser speckle reflected from an acoustically driven membrane, can act as a contactless acoustic sensor that images thousands of surface points at once. The catch is that this optical chain is a nonlinear, low-pass, spatially distributed transducer, and its raw event stream looks nothing like sound.

Rather than treat it as a demo, we treat it as an instrument: we characterize its physics and learn to invert it. A compact learned spectral decoder recovers multi-tone spectra (5–1000 Hz, up to six simultaneous tones) from events with sub-bin frequency accuracy, reaching a tone-detection F1 of 0.84 against 0.34 for a model-free peak-picker. A reference-microphone control shows that the sensor’s strong harmonic distortion comes from the optical chain, not the loudspeaker, and the decoder learns to suppress it.

We also reconstruct audible tones from unseen recordings and test transfer to real broadband sources, where a physics-aware spectral front-end classifies more data-efficiently than end-to-end networks. The result is a characterization-and-inversion method for optical acoustic sensing.