A study of speaker and noise adaptation for nonnegative matrix factorization based speech enhancement techniques

[[abstract]]本論文中，我們主要是將基於非負矩陣分解法(non-negative matrix factorization, NMF)的語音強化法探討其語者調適與雜訊調適的效能。而調適分別運用於三種NMF語音強化法，分別為全頻式、分頻式與區塊式。全頻式NMF法是直接對於語音之全頻帶的時頻圖做更新，分頻式NMF法則是分別對語音之低頻帶與高頻帶的時頻圖做更新，區塊式NMF法是先將時頻圖中各音框分別組合其時間前後之音框為一區塊、再對由這些區塊前後排列所組合成的延伸時頻圖做更新。上述之三種NMF語音強化法其語音與雜訊的NMF頻譜基底矩陣是以監督式的學習方式獲得，因此當訓練資料與測試資料不匹配時，會造成強化性能下降。本論文採用線性轉換的方式，藉由調適語句或雜訊，對於NMF頻譜矩陣加以調適，以助於降低不匹配之影響，進而提升NMF語音強化法的效能。在評估實驗中我們引用AURORA-2連續數字語料庫之部分語句，實驗結果顯示上述之調適運作得以明顯提升三種NMF語音強化法處理之語音訊號的品質，足見調適之重要與效果。[[abstract]]In this study, we focus on evaluating the adaptation performance for speech enhancement algorithms based on the nonnegative matrix factorization (NMF) technique under various distortion environments. The NMF spectral basis matrices for both speech and noise are obtained in a manner of supervised learning, and thus the performance of their associated NMF speech enhancement degrades as the speaker and/or noise characteristics are not matched for the learning and evaluation environment. An adaptation scheme is thus adopted for the NMF spectral matrices in order to alleviate the aforementioned mismatch effect and lead to superior level of speech enhancement. The experiments conducted on a subset of the Aurora-2 connected digit database show that a linear adaptation for the speech/noise spectral matrices provides the NMF-based enhancement algorithms with significant improvement in elevating the speech quality under a wide range of signal-to-noise (SNR) environments.[[note]]碩