環境雜訊變動下的語音偵測技術

Speech Detection under Non-Stationary Noise Environment

黃柏凱(Po-Kai Huang) ； 黃泰惠(Tai-Hwei Huang)

《電腦與通訊》 120期 (2007/06) Pp. 92-98

https://doi.org/10.29917/CCLTJ.200706.0015

摘要

助聽器的基本原理主要是將接收到的信號進行頻帶能量提升以補償使用者的聽力損失。由於實際環境中充滿各種雜訊，若助聽器無法判斷所接收到的信號是語音或雜訊而同樣當作語音處理的話，則使用者在配戴期間將一直被迫接受雜訊而易造成疲倦。為改進助聽器在實際環境中的使用舒適度，本研究發展適用於非穩態雜訊環境下的語音偵測技術，期望針對2分貝／秒和3分貝／秒變化的雜訊環境下可進行語音偵測。使用的語音偵測方法是擷取Cohen的改良式遞迴平均最小值預估法所估測的語音存在機率值做為語音偵測的判斷依據。在實驗過程裡，經過訓練階段找尋理想的語音存在臨界值，並且在測試階段以此理想的語音存在臨界值進行測試，分別得到平均第一類型錯誤率（語音偵測成雜訊）為18.28%和第二類型錯誤率（雜訊偵測成語音）為17.89%。

關鍵字

助聽器 ； 語音偵測 ； 非穩定雜訊

並列摘要

The fundamental principle of hearing aid is to compensate the hearing loss by providing an extra energy gain for each spectral band of received signal. Since the practical environment is always full of noises, the advanced hearing aids have to have the ability to tell speech from noise and to apply a different processing scheme for either case. Otherwise, the hearing aid users will suffer from over emphasized noise all the time. To deal with this problem, we develop the technique of voice activity detection in the non-stationary noise environment, especially for the noise energy changes at either 2dB/sec or 3dB/sec. The voice activity detection method takes advantage of the speech presence probability obtained from the Cohen's improved minima controlled recursive average method. In the experiment, we search the optimal threshold of the speech presence probability in the training phase, and we obtain an averaged Type Ⅰ error rate (noise detected to be speech) of 18.28% and an averaged Type Ⅱ error rate (speech detected to be noise) of 17.89% in the testing phase.

並列關鍵字

Hearing Aid ； Voice Activity Detection ； Non-Stationary Noise