適用於全耳道式數位助聽器之低耗電雜訊及回授消除系統設計

Abstract

由於積體電路製程與訊號處理方面的進步下數位助聽器已經具有進一步提升效能及使用者經驗的能力。然而，因為目前助聽器仍受限於有限的電池容量，故要在助聽器上實現複雜的演算法仍然是非常的困難；目前為止，最有效的方法仍為透過有效率的低耗電演算法，架構與電路設計達成。在本篇論文中，我們提出助聽器中兩個非常重要的功能，雜訊消除(noise reduction)與回授音消除(feedback cancellation)的低功率解決方案。 在雜訊消除方面，我們基於符合人耳特性的聽覺式分頻(perceptual frequency decomposition)概念，提出三種設計。在第一種設計中，我們提出一個混和式的分頻方式(mixed decomposition)，搭配簡單但有效率的頻譜相減法(spectral subtraction)及語音偵測(voice activity detection, VAD)，來達成超低耗電的雜訊消除；本設計透過0.18μm標準元件庫製程，在1.0V之下耗電僅為0.65μW。本設計可提供不錯的SNR提升，然而，因為其採用較簡單的雜訊消除方法，故無法提供較佳的聲音品質。為解決此問題，在第二個雜訊消除設計中，我們基於助聽器的ANSI S1.11 1/3-octave分頻濾波器組(filter bank)，提出時間軸多頻帶頻譜相減法(sample based multiband spectral subtraction)及時間軸焗語音偵測(sample based entropy VAD)；此方法中，我們提出預處理(preprocessing)以及其他低功率的最佳化來達成助聽器耗電以及即時(real time)的需求。本演算法可達成不錯的SNR以及音質方面的提升，基於90nm高閾值(high threshold voltage)標準元件庫製程的實現則顯示其可在0.6V之下達成約83.7μW的耗電。然而，此雜訊消除方法仍受限於VAD本身在低訊雜比下的高誤差率影響，為解決此問題，第三個雜訊消除設計中，我們基於人類語音中最強健的音高(pitch)部分，設計出基於音高的語音偵測器，實驗顯示本偵測器在低SNR甚至變動(SNR或雜訊性質)的環境中，仍能達到穩定的高性能，基於音高的概念，我們進一步設計了雜訊消除方法，實驗顯示，本雜訊消除在變動環境下亦可以提供穩定的效能，SNR約可提升4dB，65nm高閾值標準元件庫製程的晶片實現則顯示其在0.5V的運作下耗電量約為55.52μW。 在回授音消除的方面，我們基於音高處裡的概念，透過音高的資訊設計出語音共振峰預估的方法，此方法可以有效率的估測語音能量的分布，並用來輔助可適性回授消除演算法(adaptive feedback cancellation)的係數更新，降低語音信號對回授消除的影響，維持穩定的助聽器增益及音質；相對於傳統的做法，本設計的運算複雜度可以大幅降低約五個數量級，但仍可提供接近的音質。針對未來的助聽器系統晶片(SOC)發展，本論文亦提出了基於音高處理的處理器架構，可以提升助聽器的效能並降低耗電。

With the advanced digital technology and signal processing, digital hearing aids have more potential to provide good performance to improve user usage experience. However, these sophisticated signal processing algorithms are still hard to be integrated due to the limitation of battery size and capacity, which demands efficient low power algorithm, architecture and circuit design. Thus, this dissertation proposes low power designs for two fundamental blocks of hearings aids: noise reduction (NR) and feedback cancellation (FC). The proposed NR designs are based on perceptual decomposition for efficient processing. The first NR design adopts a mixed frequency decomposition in conjunction with an efficient spectral subtraction and VAD (voice activity detection) for ultra low power noise suppression. The design can achieve about 4dB SNR improvement in low SNR environment and only consumes 0.65μW at 1.0V operation using 0.18μm process. However, this design adopts a simple scheme for NR, thus not providing good perceptual performance. To solve this problem, the second NR proposes an efficient multiband spectral subtraction design by using sample based processing, data preprocessing scheme and other sophisticated strategies to meet low power and low latency requirement. This design can achieve robust sound quality improvement in terms of SNR, PESQ and composite measure with 83.7μW at 0.6V operation with 90nm HVT (high VT) standard cell library. The performance of the second design is limited by the accuracy of entropy VAD in low SNR and nonstationary environment. To solve this problem, the third design proposes an efficient pitch based VAD for robust voice detection to assist noise suppression. This VAD has an efficient structure and is robust even in nonstationary environment. Based on this VAD, the noise suppression can provide 4dB SNR improvement with 55.52μW at 0.5V operation with 0.65μm high VT process. The pitch based processing is further applied to FC design which uses pitch results to estimate speech formant to enhance the robustness and the sound quality of adaptive feedback cancellation (AFC). The proposed AFC design can achieve similar added stable gain (ASG) and PESQ but with five orders complexity reduction compared to conventional designs. Based on the pitch based information, this dissertation also proposes an efficient pitch based processor for further system development.