適用於數位助聽器之快速傅立葉轉換的可調式步階係數與探測訊號之回授消除技術以及系統震盪偵測設計

Abstract

由於製程的進步以及數位訊號處理方面的廣泛，數位助聽器已成為現在助聽器主流。然而，由於目前助聽器仍受限於電池容量太小，以及為了避免患者對語音和眼睛所看到的影像產生誤差，對訊號的即時處理是必須的，因此要實現複雜的演算法是比較困難的。而到目前為止最有效的解決方式是透過有效率的低功耗演算法，架構以及電路設計去完成。對於使用者來說，最在乎的不僅僅是對聲音的放大和噪音的消除，還有一重點是對於刺耳尖銳聲的預防和抑制，由於尖銳聲對患者來說是不舒服也難以忍受的，因此這也大大地降低了患者配戴助聽器的意願。尖銳聲的產生，即是回授訊號沒被消除且過度放大的結果，為了解決這一問題，我們提出了一個快速傅立葉轉換的調式步階係數與探測訊號之回授消除技術及系統震盪偵測設計來防止和抑制尖銳聲的產生。 在我們所提出的回授消除技術及系統震盪偵測演算法中，最重要的資訊皆來自於音高式語音偵測器。音高式語音偵測器是利用語音裡音高以及相對應的和諧音和子音起始的特性去偵測語音。為了更進一步改良我們的準確度，我們把兩種方法結合來偵測音高以及相對應的和諧音。 在回授消除部分設計出一個語音共振峰預估的方法。這個方法是利用音高的資訊去預估語音能量的分布，並用來輔助可適性回授消除演算法的係數更新。此外，針對能量變化差異較大的輸入訊號環境，我們設計了一套機制，來調整可適性回授消除演算法的步階係數或是加入不同大小的探測訊號到系統中，藉此來調整可適性回授消除演算法的收斂速度，來維持穩定的助聽器增益及音質。 對於系統震盪偵測以及尖銳聲音的消除，我們利用經由快速傅立葉轉換後的訊號，在各頻率區間做能量監視，來偵測出那些能量區間是有可能產生尖銳聲音，並且藉由音高式語音偵測器中所偵測出的音高以及相對應的和諧音做保護，以防止誤刪除語音訊號，之後經由尖銳聲消除機制來判定是否消除或是抑制其能量。

With the advancement of IC technology and digital signal processing, digital hearing aids have been the main trend of hearing aids. However, it is difficult to implement a complicated algorithm due to the limitation of battery size and capacity. Moreover, the real-time processing is necessary to match the speech with the image which the patients perceive. Therefore, it is difficult to implement a very complicated algorithm. The effective way of solving this problem is to design an efficient low power algorithm, architecture and circuit. For hearing aid users, they care not only about speech enhancement and noise reduction, but also about the prevention and suppression the howling sound. The howling sound makes the patients uncomfortable and is therefore unacceptable. Hence, it will reduce willingness to use hearing aids. The howling sound is the result of feedback signals that are not cancelled completely. In order to solve these problem, we propose an FFT based feedback cancellation and oscillation detector. The most important novelty of our proposed feedback cancellation and oscillation detector is based on the pitch based voice activity detection. The pitch based VAD utilizes the pitch and its harmonics and onset characteristics of speech to detect speech activity. In order to improve the VAD accuracy, two kinds of methods are combined for pitch and harmonic detection. The design of feedback cancellation is based on FFT decomposition and a decorrelation filter coefficient update mechanism is proposed. The decorrelation filter coefficient update utilizes the pitch information to estimate speech formant to enhance the robustness and the sound quality of adaptive feedback cancellation (AFC). In addition, we design a mechanism to adjust the step size of the adaptive feedback algorithm or inject the probe noise with different magnitude for large magnitude variation input signal. By this mechanism, we can adjust the convergence rate of adaptive feedback algorithm to achieve a higher stable gain and sound quality. The PFE-AFC achieves the comparable PESQ and ASG compared with PEM-AFC in previous work. Moreover, the new version of PFE-AFC which we proposed can improve the PESQ by about 0.03 score and the minimum of ASG about 4dB. In high magnitude variation of input signal, the new version of PFE-AFC can provide more stable ASG and 7dB enhancement. The design of oscillation detector and howling suppression utilizes the signal after FFT to monitor the magnitude in different frequency region. Then, detect the potential instable frequency region where howling may occur. Besides, we protect the pitch and its harmonics by pitch based voice activity detection to prevent the howling suppression from canceling the speech information. Then, the howling suppression decides whether to reduce or cancel the potentially instable frequency.