感知式機器學習應用於助聽器自主降噪及方向性調整技術

Self-Administered Noise Reduction and Directivity Adjustments of Hearing Aids Using Perception-Based Machine Learning

王彥博(Yen-Po Wang) ； 楊謦維(Ching-Wei Yang) ； 張弘岳(Hung-Yue Chang)

《應用聲學與振動學刊》 14卷1&2期 (2022/12) Pp. 29-39

https://doi.org/10.6282/JASV.202212_14(1_2).0004

摘要

為了在沒有專業聽檢室或聽力師的情況下進行助聽器個人化調整，我們期望讓使用者可以自行於所處環境下得到個人化的降噪及方向性參數，因此使用改良式高斯過程機器學習演算法，以預測個人化參數，並結合語音清晰度及語音品質相關的指標，分別是語音清晰度品質（speech intelligibility quality, SIQ）及環境噪音品質（environment noise quality, ENQ），並以全局語音品質（global speech quality, GSQ）作為個人化調整，多種助聽器參數交互作用下，也能找出不同環境下，針對個人注重語音清晰度、語音品質或全局語音品質的個人化參數配置。二維實驗結果顯示預測的最佳配置與初始配置相比，SIQ、ENQ、GSQ三者的得分依序提升0.352、0.409、0.238，考量不同指標下的四維實驗，結果則顯示在最佳化配置下，SIQ、ENQ、GSQ三者的得分相較於初始配置，依序提升0.455、0.757、0.537，證明此研究確實能做到自主的降噪及方向性個人化調整，同時考量到使用者注重的不同指標。

關鍵字

自我選配 ； 個人化 ； 機器學習 ； 語音清晰度 ； 語音品質

並列摘要

As the population of hearing-impaired increases in recent decades, the hearing-aid personalization without a professional audiometry or an audiologist's assistance becomes feasible using a self-fitting hearing aid. A modified Gaussian process (GP) machine learning algorithm is used to predict the personalized parameters of noise reduction and directivity based on individual's target of speech intelligibility and speech quality-related indicators, namely speech intelligibility quality (SIQ), environment noise quality (ENQ), and global speech quality (GSQ). Under the interaction of various hearing aid parameters, it is also possible to find individual parameter configurations that focus on speech intelligibility, speech quality or global speech quality in different environments. The two-dimensional experimental results show that the predicted optimal configuration improved the scores of SIQ, ENQ, and GSQ by 0.352, 0.409, and 0.238, respectively in comparison with the initial configuration. The four-dimensional experimental results show that compared with the initial configuration, the predicted optimal configuration improved the scores of SIQ, ENQ, and GSQ by 0.455, 0.757, and 0.537, respectively. This study demonstrated that the GP machine learning can indeed achieve personalized adjustments of noise reduction and directional parameters, and consider different user preferences of SIQ, ENQ, and GSQ at the same time.

並列關鍵字

Self-Fitting ； Personalization ； Machine Learning ； Speech Intelligibility ； Speech Quality