有關響度與分頻解析度之聽損耳蝸模型的建立

Abstract

為了能夠即時驗證助聽器演算法的效果，我們必頇有一套能夠模擬聽損病患耳蝸的系統才行；因此我們參考Moore 團隊所提出的，模擬最小可聽水平提升、響度聚集、以及分頻解析度降低的耳蝸模型，其中有關響度的變化是以濾波器組的方式處理，將訊號分頻之後把能量N 次方，放大相鄰時間點能量上的差距；而分頻解析度降低則是以頻譜模糊化來取代，將基底膜接收聲音的機制用激發模式表示，並將其矩陣化後以反矩陣的方式來完成，最後在短時傅利葉轉換的架構下進行。然而這樣子的做法有幾個缺點：1.將能量N 次方只能決定響度聚集的程度，無法決定最小可聽水平的值。2.短時傅利葉轉換中保留原相位頻譜會造成模擬程度不匹配，此外重疊的濾波器組解析度被傅利葉轉換所限制。3.需經過兩次短時傅利葉轉換與一次濾波器組的架構，計算複雜度高。因此，我們首先將響度變化的機制以分段的線性公式一點一點的計算；接著我們將變寬的聽覺濾波器用正常濾波器的線性組合組合而成，計算出原始的濾波器對所有變寬濾波器的貢獻，並將白高斯雜訊濾波過過後取出載波，和經過模糊化的能量相乘後合成為聲音訊號，如此整個系統便可在濾波器組的架構下完成。最後我們設計聽力測驗，證明我們的模型與Moore 團隊的模型會有相似的結果，也符合聽損患者回報出的一些聽損現象。

In order to verify the effectiveness of algorithms developed for hearing aids, developing a system that simulates the cochlear of the hearing impaired is in great demand. Thus, we investigate the model developed by Moore’s group in addressing threshold elevation, loudness recruitment, and reduced frequency selectivity of hearing impaired. The simulation about loudness was done by extracting the sub-band signal and powering the instantaneous magnitude according to the degree of recruitment. To simulate the frequency smearing, matrix multiplications of filter-bank coefficients of the normal/impaired cochlear were carried out. However, there are some drawbacks in Moore’s model. First, the minimum thresholds can not be set alone. Second, preserving the phase spectrum would mitigate the desired degree of magnitude smearing. Also, the spacing of the critical bands is restricted by FFT. Third, the computational cost is very high. To overcome these drawbacks, we (1) use a piecewise formula to decide the loudness sample by sample; (2) model the widened filter by a linear combination of normal filters and compute contributions from normal filters to each widened filter. Besides, we use white Gaussian noise to generate the band-passed carrier and synthesize the signal. Our approaches can be done in the same filter-bank architecture to reduce computational load. Last, listening tests are carried out to verify our model. Results show that our model not only acts like Moore’s model, but also exhibits characteristics reported by the hearing impaired.