助聽器系統的時頻分解、降噪、與適應性回授消除之分析與設計

Abstract

由於近年無線通訊的快速進步，加上新型態的應用需求不斷地被挖掘，可以預期目前傳統之數位助聽器，會隨著當下雲端計算與物聯網之趨勢有所改變，同時也能進一步改進助聽器的使用者經驗。我們預測未來助聽器晶片系統，必須提供更多的訊號處理，來滿足新型態應用之需求，在有限的電池容量與計算能力下，要同時滿足此需求，是需要由系統應用層級、信號處理演算法、和硬體實現架構三者，系統化的討論後，才能提出一套完整的設計方法論，最終再由應用需求與條件，選擇最佳的助聽器系統。本論文之核心目的，是將系統化的討論助聽器系統中，關於時頻分解、降噪、與適應性回授消除，這三大獨立主軸與之間的系統演算法探討，與硬體加速器設計。 第一個主題將探討以ANSI S1.11濾波器組的規格為基礎，在延遲、複雜度、子頻帶資料量、合成輸出漣波與處方簽補償誤差這五項目進行討論，並設計一套濾波組設計流程，與如何實現一個高效率的多率架構。 第二個主題是一個給多率有限長度濾波器的一種高效率多相分解，我們進一步發現，此多相分解具有規律性，且在分數式的取樣轉換器上，比任何一種多相分解，都有約1.5至2倍的乘法運算量減少，是可以應用於第一個主題提出的多率濾波器組實作，我們由此多相分解的規律性中，找出最小運算單元，將之實作成硬體架構，同時也完成一個對應的執行程序電腦程式，使其能應用於任何一種多率的有限長度濾波器硬體實現上。 接著第三個主題，提出一個完整的單耳助聽器系統，其包含音高式降噪、適應性共振頻譜估測回授消除，與回音抑制三個主要子系統，由於單耳系統架構下的限制，造成兩個無法解決的問題，一是無法有效對付多人聲下的背景雜訊消除，二是降噪與回授消除演算法之間衝突的影響，因此單耳助聽器系統只針對非人聲背景雜訊作消除，並利用系統的情境控制機制，協調開啟降噪與適應性演算法的時機，避免互相衝突，減少系統效能損失，然而，降噪與回授消除中的適應性演算法無法同時開啟的情況下，我們需要回音抑制子系統，來避免當回音產生時，盡可能避免演算法效能損失，並對回音進行立即消除；另外，我們於回授消除演算法中，提出一種雜訊饋入的方法，來穩定整體系統的穩定度，最後我們對此系統進行不同種背景雜訊與使用者情境的模擬，並分析其結果。 第四個主題是針對未來雙耳助聽器之降噪應用來討論，以不同規格之濾波器組，與雙耳降噪演算法之間，作演算法效能的探討，我們依據演算法效能分析結果，使用多率架構來降低運算量與資料量，並進一步將高計算需求之部分，設計加速硬體，驗證其實作結果。 最後，本論文完成之研究探討，可以在助聽器的系統演算法設計，與異質整合的晶片系統中有所貢獻。

As the rapid development of wireless communication technology and emerging application-specific demands, we expect that the conventional digital hearing-aids will evolve with the current cloud computing and internet of things, and thereby the user experiences can be significantly improved. We predict that this future hearing-aid system must provide more signal processing capability to satisfy the new type of demands. Under the limited battery capacity and computing power, it needs a systematic discussion from application-specific system level aspects, signal processing algorithms, and hardware implement architectures to meet these requirements. A complete design methodology can be invented, then we can chose the most beneficial hearing-aids according to the application specific demands and conditions. The core objectives of this dissertation is to systematically discuss the systems' algorithms and hardware accelerator designs surrounding on the topics and interactions of time-frequency decomposition, noise reduction, adaptive feedback cancellation. Based on the ANSI S1.11 filter specification, the first topic is to discuss with the delay, complexity, subband rate reduction, synthesized output ripples, and prescription matching errors. A complete filter bank design flow is presented, and it explains to implement an high efficiency multirate architecture. The second topic is a high efficient polyphase decomposition for multirate linear-phase FIRs. We have discovered that this decomposition algorithm have some excellent regularities. If a fractional sampling rate conversion is adopted this method, it has 1.5 $\sim$2 times multiply savings than others. Therefore, we believe that it can be applicable for the hardware implementation of first topic. According to the regularity of polyphase decomposition algorithm, we have found the basic computing elements and turned into specialized hardware accelerator. Meanwhile, we also have developed a computer program that can automatic generate the computing sequences for the specialized processor. The third topic proposed a complete monaural hearings that includes pitch-based noise reduction, formant estimation based adaptive feedback cancellation, and howling suppression sub-system. Due to the restrictions on monaural hearing-aids, it results in two unsolvable problems. One is not effective to deal with multi-taller noise, and another is the conflicts between noise reduction and feedback cancellation. Our monaural hearing-aids are designed to remove the unvoiced background noises, and we assume that a scene detection mechanism is added to the system to coordinate the noise reduction and adaptive feedback cancellation without great performance degradation. However, the adaptation of feedback cancellation will be not always enabled under this circumstance. Therefore, we need howling suppression to minimize the performance degradation from howling removal. In addition, we also proposed a noise injection method to stabilize the overall system. Finally, we have the simulations with different noise types and user scenarios to verify and analyze the proposed system. The last topic focuses on the future binaural noise reduction hearing-aids. We adopt the variant ANSI-like filter banks on binaural noise reduction to investigate the algorithmic performance. We referred the performance evaluation results and used multirate techniques to save the subband data rate and computational complexity. Finally, we realize the high computing demand parts into hardware accelerators and verify the implementation results. In the end of dissertation, we believe that our research would make a valuable contribution to the algorithmic system architect and integrated heterogeneous chip systems for hearing-aids.