語音增強基於小腦模型控制器

本文提出了一個小腦模型控制器（Cerebellar Model Articulation Controller, CMAC）應用於語音增強系統（Speech Enhancement System），所提出的CMAC使用歸一化梯度下降法（Normalized Gradient Descent Method）增加CMAC參數的自適應學習速度，具有比傳統類神經網路方法更快的學習速度、體積小且良好的泛化，因此更適合做高速的訊號處理。實驗方面，使用CMAC與MMSE做比較，為了比較性能，我們用了三種語音評估方法來做CMAC消除雜音及MMSE消除雜音後的數值比較，分別為（Perceptual Evaluation of Speech Quality, PESQ）、（Segmental Signal-to-Noise Ratio, SSNR）以及（Speech Distortion Index, SDI）。由實驗結果可知，在三種評估方法，CMAC皆能達到較佳的結果。

Traditionally, cerebellar model articulation controller (CMAC) is used in motor control, inverted pendulum robot, and nonlinear channel equalization. In this study, we investigate the capability of CMAC for speech enhancement. We construct a CMAC-based supervised speech enhancement system, which includes offline and online phases. In the offline phase, a paired noisy-clean speech dataset is prepared and used to train the parameters in a CMAC model. In the online phase, the trained CMAC model transforms the input noisy speech signals to enhanced speech signals with reduced noise components. To test the CMAC-based speech enhancement system, this study adopted three speech objective evaluation metrics, including perceptual evaluation of speech quality (PESQ), segmental signal-to-noise ratio (SSNR) and speech distortion index (SDI). A well-known traditional speech enhancement approach, minimum mean-square-error (MMSE) algorithm, was also tested performance for comparison. Experimental results demonstrated that CMAC provides superior performances to the MMSE method for all of the three objective evaluation metrics.