DAT-CFTNet: Speech Enhancement for Cochlear Implant Recipients using Attention-based Dual-Path Recurrent Neural Network

A team of researchers led by Nursadul Mamun and John H.L. Hansen has introduced DAT-CFTNet, a new AI model designed to significantly improve speech clarity for cochlear implant (CI) recipients. The system is built by integrating a novel dual-path attention recurrent neural network (DAT-RNN) with an enhanced complex-valued frequency transformation network (CFTNet). This architecture is directly inspired by the human auditory system's ability to focus on key speech elements, allowing the AI to precisely differentiate between speech and noise within the time-frequency regions of an audio spectrogram. By optimizing both local and global context processing, it tackles a core challenge for CI users, who often experience severely limited hearing restoration in noisy settings.

The model demonstrates consistent performance improvements over existing benchmarks like the standard CFTNet and DCCRN models, particularly in metrics for speech intelligibility and quality. Crucially, it excels at suppressing non-stationary background noise while avoiding the 'musical artifacts'—unnatural, tonal noises—that plague many traditional speech enhancement methods. This targeted enhancement is vital, as studies show CI listeners can have greater than 10% limitations in time-frequency resolution. The researchers' implementation will be made publicly available, and the paper is slated for presentation at the 2026 IEEE ICASSP conference, marking a promising step toward more effective assistive hearing technology.