Runs neural audio inference directly on the Neural Engine, bypassing CoreML: about 157 µs per 8-voice buffer, roughly 79× real-time headroom, with no CPU cores spent on inference.
Cite this entry
@misc{tbc_ane_synth,
title = {A real-time neural synthesizer on the Apple Neural Engine},
author = {{TBC Research}},
note = {F-006, catalogue version 1},
howpublished = {\url{https://tbcresearch.org/research/ane-synth/}}
}