經驗模態分解能根據訊號本身的特性提取具有物理意義的本質模態函數。因此在處理複雜的語音訊號時，經驗模態分解具有顯著的優勢。然而經驗模態分解提取的本質模態函數經常出現模態混合現象，導致其物理意義受損。因此許多擾動輔助經驗模態分解被提出，例如總體經驗模態分解與均勻相位經驗模態分解，來改善此問題。儘管擾動輔助經驗模態分解在處理非線性及非平穩訊號方面具有優勢，但其高運算量與記憶體需求對資源有限的嵌入式系統構成挑戰。在本研究將評估擾動輔助經驗模態分解演算法在嵌入式穿戴裝置中常用的微控制器中即時處理語音訊號的可行性，並在符合即時語音處理限制的條件下進行最佳化以降低計算時間、記憶體需求與系統延遲。經過最佳化，研究結果顯示，使用即時自適應擾動輔助經驗模態分解演算法進行語音除噪時，對於採樣頻率為16k Hz的語音訊號，計算負載比達54%、音訊延遲為0.03秒，記憶體需求為18.75 KB；而進行語音特徵處理時，對於8k Hz的語音訊號，計算負載比為59.67%、音訊延遲為0.1697秒，記憶體需求為68.75 KB。雖然在微控制器上實現以擾動輔助經驗模態分解進行即時語音處理具有可行性，但計算量與音訊延遲依舊是很大的問題。;Empirical Mode Decomposition (EMD) extracts intrinsic mode functions (IMFs) with physical significance based on the signal′s characteristics, making it advantageous for complex speech signal processing. However, EMD often suffers from mode mixing, which undermines its physical interpretation. To address this, disturbance-assisted EMD (DA-EMD) methods like Ensemble EMD (EEMD) and Uniform Phase EMD (UPEMD) have been proposed. While DA-EMD excels in handling nonlinear, non-stationary signals, its high computational load and memory requirements pose challenges for resource-limited embedded systems. This study evaluates the feasibility of implementing DA-EMD for real-time speech processing on microcontrollers (MCUs) in embedded wearable devices, optimizing it to reduce computation time, memory usage, and system latency. The optimized Real-Time UPEMDA algorithm demonstrated a 54% computational load, 0.03-second audio delay, and 18.75 KB memory usage for 16kHz speech denoising. For 8kHz speech feature extraction, the computational load was 59.67%, with a 0.1697-second delay and 68.75 KB memory usage.
