| 摘要: | 本論文目的是發展針對風聲噪音下的風聲除噪系統。當風聲吹過麥克風時，會產生巨大的噪音，影響收音品質，降低訊雜比。此系統包括兩個部分：一是計算輸入之梅爾頻譜倒參數 (Mel Frequency Cepstrum Coefficients)，再利用機器學習中的分類樹來自動判斷當前風切聲噪音存在與否，再改變適應性濾波器的參數進行除噪；二是計算輸入的短時譜熵 (Short Time Entropy) 來偵測麥克風接收到的語音活動，若偵測不到語音，則消除噪音，以增加輸出的聽覺舒適度和理解度。為了驗證系統能有效去除不同的風聲，使用了真實的和模擬的風聲作為噪音來源。並與目前對於噪音消除表現較佳的演算法最小控制遞迴平均法 (Minima Controlled Recursive Averaging, MCRA) 和正反向最小控制遞迴平均法 (MCRA-FB) 比較。輸入訊號從訊雜比為10dB到-10dB。使用語音品質客觀評量方法 (Perceptual Evaluation of Speech Quality, PESQ) 評量最後結果。
首先利用MATLAB (The MathWorks, Natick, Massachusetts, USA)之軟體模擬本驗算法，在訊雜比為0dB 時，PESQ評分可以比未除噪提高 0.35，MCRA-FB只能較未除噪提高0.05。風噪聲判斷準確率為93%，語句判斷率準確率為96%。本研究也實現在TMS320C6713開發板 (Texas Instruments, Dallas, Texas, USA)。當使用line in輸入在高訊雜比 (6dB) 時，PESQ評分可以比未除噪提高 0.25，而MCRA演算法應用在此不能有PESQ分數的提升。若是使用適應性指向麥克風收音，在低訊雜比 (-10dB) 時，可以比未除噪提高0.35的PESQ分數，MCRA法只能提高0.2。本研究在電腦模擬和硬體實現上，可以有效去除風聲噪音，語句之間的風聲噪音也可以完全刪除，較目前常用的演算法如MCRA方法等等表現優良。
;As we have known in our daily life, a great noise on the microphone would be produced and signal-to-noise ratio (SNR) of the perceived speech and its quality would be lowered when wind passes through the microphone. The purpose of this study was to develop an adaptive wind noise reduction system. Our system has two parts: firstly we applied the decision tree machine learning algorithm to detect existence of wind noise with the mel frequency cepstrum coefficients (MFCC) used as input features, and parameters of adaptive filter would be changed to reduce the wind noise. Then we calculated the input short time entropy to detect the voice activity in order to make the output speech signal more comfortable and intelligible. This approach would reduce the wind noise if it detected the input signals with no speech activity. To verify if our system could reduce different wind noise properly, we applied real and simulated wind noise as the noise sources with SNR set from 10 to -10dB, and compared our results with two common noise reduction algorithms: minima controlled recursive averaging (MCRA) and Forward-Backward MCRA (MCRA-FB). Then the objective perceptual evaluation of speech quality (PESQ) approach was used to evaluate the quality of the results.
In this study, the MATLAB (The MathWorks, Natick, Massachusetts, USA) program was first used to implement the wind noise reduction system. Our results showed that the PESQ score was increased by 0.35 when compared to the original signal with 0dB SNR real wind noise signal while MCRA-FB algorithm could only increase by 0.05. At the same time, the speech hit rate was 96%, and the accuracy of the wind noise detection rate is 93%. We further implemented the wind noise reduction system on the DSP starter kit (DSK), TMS320C6713 (Texas Instruments, Dallas, Texas, USA) and compared to the results of MCRA. Our results indicated when the line in was used as the signal input, the PESQ score could be increased by 0.25 at high SNR (6dB) signal while the results of MCRA algorithm could not improve the PESQ score. However, when the adaptive directional microphone (ADM) was used as the signal input, the PESQ score of our result was 0.35 higher than that of the original (no noise reduction) system at low SNR (-10dB) signal while the result of MCRA algorithm only improved by 0.2. These results show that our wind noise reduction system could reduce the wind noise properly and achieve better performance than the MCRA algorithm. |
