6.3.2 - A Dual-Resolution Binary Neural Network Inference Engine for Fully Implantable Cochlear Implant Speech Processing

M. Doğan; H. Uluşan; H. Külah

6.3.2 - A Dual-Resolution Binary Neural Network Inference Engine for Fully Implantable Cochlear Implant Speech Processing

Event: EUROSENSORS 2026
2026-09-06 - 2026-09-09
Zurich
Band: Lectures
Chapter: Signal Analysis, Machine Learning and Artificial Intelligence for Sensing
Author(s): M. Doğan, H. Uluşan, H. Külah - Metu,Ankara (Turkey)
Pages: 211 - 212
DOI: 10.5162/eurosensors2026/6.3.2
ISBN: 978-3-910600-12-6
Price: free

Abstract

A dual-resolution Binary Neural Network inference engine replaces conventional filterbanks in cochlear implant speech processing. A 1 kHz split applies a 256-sample window (75% overlap) to low-frequency channels and a 128-sample window (50% overlap) to high-frequency channels, with a shared 64-sample hop. The architecture matches fixed N=256 intelligibility (STOI 0.826 vs 0.830) while halving high-frequency channel latency and reducing L1 inference cost and size by ~25%, at <2 mW inference power. The approach can be applied to edge AI sensor SoCs requiring low-power spectral analysis.