The Evolution of System on Chip Integrated Circuits for Hearing-Aid Signal Processing

Abstract

Integrated circuits used in modern, digital hearing aids are mixed-signal devices that combine analog and digital circuits on the same die. Due to the extreme size and power constraints encountered in hearing aids, the assignment of die area for such systems-on-chip (SOCs) must be carefully considered. The relative allocation between analog circuits, digital gates and on-chip memory has been relatively constant for traditional DSP-based systems. The precision and computational requirements for artificial intelligence (AI) based algorithms, however, are very different from traditional DSP. This is driving changes in die area allocation, starting with some recently launched devices. Despite these changes, projections based on recent growth trends show that future SOCs will only be able to support the smallest neural networks currently proposed for hearing aids. Minimizing data precision and pruning small-valued weights are shown to be essential to deploy deep-learning algorithms on hearing aid SOCs.