Pathak Nishank, Panahi Issa, Devineni P, Briggs Richard
University of Texas at Dallas, Richardson, TX 75080, USA.
Annu Int Conf IEEE Eng Med Biol Soc. 2009;2009:6950-3. doi: 10.1109/IEMBS.2009.5333749.
Performance of two Adaptive (nLMS and Normalized Sign-error LMS) and a single channel (LogMMSE) speech enhancement algorithms are tested on a floating point DSP to reveal their effectiveness in enhancing speech corrupted in noisy MRI environment with very low SNR. The purpose of experiments is to reduce the fatigue of the listener by eliminating the strong MRI noise. The experiments use actual data set collected from a 3-Tesla MRI machine. Results of the experiments and performance of the speech enhancement system are presented in this paper. The speech enhancement system is automated. Our experiments reveal that after enhancement of the speech signal using Sign-Error LMS, the residual noise shows characteristics of white noise in contrast to the residual noise of the other algorithms which is more structured. It is also shown that the Sign-Error LMS offers fast convergence in comparison to the other two methods.
在浮点数字信号处理器上测试了两种自适应算法（归一化最小均方算法和符号误差归一化最小均方算法）以及一种单通道算法（对数最小均方误差算法）的性能，以揭示它们在增强信噪比极低的噪声MRI环境中受损语音方面的有效性。实验目的是通过消除强烈的MRI噪声来减轻听众的疲劳。实验使用了从3特斯拉MRI机器收集的实际数据集。本文给出了实验结果和语音增强系统的性能。语音增强系统是自动化的。我们的实验表明，使用符号误差最小均方算法增强语音信号后，残余噪声呈现出白噪声的特征，而其他算法的残余噪声更具结构性。还表明，与其他两种方法相比，符号误差最小均方算法收敛速度更快。
