人工電子耳改變了數十萬無法有效使用助聽器的聽損者，然而，其聲音品質至今在聲調語言、噪音環境中的語音以及音樂等方面仍有限制。本計畫延續上一年度之計畫，除了繼續探討三個以聽覺生理為基礎的聲音處理方法，並進一步發展基於深度學習的聲音處理策略，以了解它們對於中文語音辨識的效果。有別於大多數電子耳和類神經網路相關的研究，本計畫提出的新穎的聲音處理策略，除了使用類神經網路訓練而產生，並將整合以深度學習之極限學習機ELM為基礎的除回響機制。由上一個計畫的初步結果顯示，增強包絡訊號的EE策略有機會增強中文摩擦音及塞擦音的電極刺激。在本計畫的第一年，我們將在現有實驗平台上完成策略的實作，並進行客觀的語音效果評估和主觀的電子耳模擬實驗。而在第二年，我們將進行本年度及上一年度計畫的臨床實驗，將處理過的訊號從電腦輸出至Cochlear公司提供的NIC 4裝置和研究用處理器，以直接串流到使用者體內的植入裝置，並評估實際的電刺激效果。透過實驗結果比較不同電子耳處理策略對於中文聲調和語音辨識在噪音與回響環境中的表現，以期有助於改善人工電子耳的聲音品質並促進相關領域的研究發展。 ;Cochlear implants (CI) have significantly changed the lives of hundreds of thousands hearing impaired people who could not use hearing aids effectively. However, the sound quality of the cochlear implant today is still limited in the performance of tonal languages, speech in noisy conditions and music perception. The proposed project is going to not only continue the previous project on three auditory physiology based sound processing methods, but also develop sound processing strategies based on the artificial neural network (ANN). In contrast to the directions of most researches on the CI and the ANN, this project suggests an innovative ANN strategy, which is generated by training with the ANN as well as integrating the dereverberation mechanism based on the extreme learning machine (ELM) of deep learning. The results from the previous project show that the envelope enhancement strategy may increase the electrode stimulation for Mandarin fricatives and affricatives. In the first year of this project, sound processing strategies will be implemented on the current experiment platform, which will be used for objective speech performance evaluation and subjective CI simulations. In the second year, clinical experiments of this project and the previous-year project will be carried out. Processed signals will be directly streamed from a computer into each CI recipient’s internal implant via the NIC (Nucleus implant communicator) 4 device and a research processor provided by Cochlear Ltd. Hence the actual electric stimulation results can be evaluated. Based on the outcomes these experiments, the performance of different CI strategies for Mandarin lexical tone identification and speech recognition is going to be compared in order to improve the sound quality of CIs and to enhance the research and development in the related fields.
