本研究係針對高規格助聽輔具的設計與發展，提出以時頻分析與卷積類神經網路計算為基礎之噪音抑制演法。期望能與前期計畫所開發之近似Class-2、Class-2、以及Class-0 ANSI S1.11 1/3 八度音濾波器組結合，透過多麥克風通道設計以及所提出之新穎滑動式/跳點式離散傅立葉轉換時頻分析器，能與卷積類神經網路達成噪音程分的估測達成增強語音品質之效果。本研究亦希望能改良相關研究，並於助聽輔具技術上有所貢獻。 This research is for the development and design of high-level specification of hearing aid, and it mainly focuses on developing the noise suppression algorithm based on time-frequency signal analysis and convolution neuro network (CNN) computation. Additionally, we will expect to integrate the proposed algorithms in this work with the key results of previous MOST projects, i.e., quasi-class-2, class-2, and class-0 ANSI S1.11 1/3-octave non-uniform filterbank design; By combining with multi-channel microphones and the proposed sliding/hopping DFT time-frequency computations, the CNN algorithm can effectively estimate the power spectrum density of noise to increase the quality of speech enhancement. Hope that it can improve the related approaches on technical developments of hearing aids.
