基於深度學習之降噪演算法應用於可保留聲響聽覺之人工電子耳

Abstract

聲碼器模擬是普遍被用來模擬人工耳蝸所產生的電訊號，聲碼器處理過後的信號讓正常聽力者聆聽時，可以臨摹聽損患者所感受到的聲音。我們計劃團隊正在研發一項聽覺聽力保留的高頻四電極點人工電子耳系統，為了要事先評估此系統的功能性，本論文利用了實驗室先前開發模擬聽損人士的個人化聽損模型，結合四通道的高頻聲碼器，分別模擬聽覺聽力保留的聲響聽覺和人工耳蝸信號處理的電聽覺，讓正常聽力者聽到聽覺保留之高頻聲碼模擬器模擬的聲音。透過中文聽辨度的心理聲學實驗，我們找尋到此人工耳蝸的最佳頻率分布方式。經過不管在乾淨語音或噪音背景下的實驗，結果顯示中文的聲母比韻母更加難以被聽損患者辨識，增加了此系統的電子聽覺，中文的聽辨度被顯著地提升了，尤其是聲母更為明顯。接著結合深度學習方式，對於生活環境中常見的噪音進行降噪演算法，討論在模型沒訓練過的噪音之下降噪，進而增加整體的中文聽辨度，並且用客觀評分標準與主觀聽辨度測量來驗證此模型。

Vocoder simulations are generally adopted to simulate the electrical hearing induced by the cochlear implant (CI). Our research group is developing a new four-electrode CI microsystem that induces high-frequency electrical hearing while preserving low-frequency acoustic hearing. To assess the functionality of this CI, a previously developed hearing-impaired (HI) hearing model is combined with a 4-channel vocoder in this thesis to respectively mimic the perceived acoustic hearing and electrical hearing. Psychoacoustic experiments are conducted on Mandarin speech recognition for determining spectral coverages of electrodes for this CI. Simulation results show that initial consonants of Mandarin are more difficult to recognize than final vowels of Mandarin via acoustic hearing of HI patients. After electrical hearing being induced through logarithmic-frequency distributed electrodes, speech intelligibility of HI patients is boosted for all Mandarin phonemes, especially for initial consonants. Similar results are consistently observed in clean and noisy test conditions. Next, we combine a deep neural network based noise reduction algorithm with the proposed CI system in the hope to improve the Mandarin speech intelligibility for seen and unseen noise types. Ultimately, we use objective evaluation and subjective evaluation scores to verify this model, hence, to provide the proof of concept of this combinational system.