ChladniSonify addresses a specific annoyance in audiovisual work: mappings between image and sound are usually arbitrary, and simulating plate modes is either slow or requires specialist software. The authors generate a paired dataset from Kirchhoff-Love plate theory, calibrate it against ANSYS finite-element simulation, and train a lightweight convolutional network with an attention module to recognise the thin nodal lines that distinguish one pattern from another.
The system runs end to end in Python and Max/MSP, mapping each recognised pattern to the sine frequency that would produce it. They report 99.33 per cent classification accuracy with 7.03 millisecond inference and under 50 milliseconds end to end — fast enough for live interaction.
Caveat: this is a nine-page arXiv preprint in IEEE conference format, posted in May 2026, and we have seen no peer review. Treat the numbers as the authors' own. The architecture is the interesting part, and it is reproducible.