A study of noise suppression approaches based on nonnegative matrix factorization for speech enhancement

[[abstract]]在本篇論文中，我們使用了分負矩陣分解法(nonnegative matrix factorization,NMF)技術來強化語音，藉此提升語音的品質與語音辨識的精確率。以往基於NMF之語音強化都是做於全頻帶的NMF 分解與更新，因此，我們探討相關的兩種改進技術：第一，將原始全頻帶頻譜拆解成高頻頻帶與低頻頻帶成分、個別對其做NMF分解與更新；第二，做於時間序列域上，對部份時間區塊之頻譜做NMF分解與更新。 另外，我們還將這些基於NMF 的新技術結合了傳統的語音強化法，如頻譜消去法和最小平均平方誤差法等，驗證其在提升語音品質的加成性。在實驗資料庫的選擇上，我們使用國際通用AURORA-2 連續數字語音庫，其中的語音訊號分別受到各種加成性雜訊與通道效應的影響。從實驗結果，可以證實我們探討的新方法能有效的提升語音的品質與語音辨識的精確率。[[abstract]]In this thesis, we exploit the technique of nonnegative matrix factorization (NMF) in speech enhancement, considering the sub-band and temporal patch characteristics of noisy spectrogram. We investigate the segmental NMF scheme in speech enhancement and compare it with the conventional frame-wise counterpart. Two forms of segmental NMF methods are investigated, which respectively decompose the spectrogram into temporal and spectral segmental parts, and then compensates each segment to reduce the noise effect. We evaluate these NMF-based methods in a subset of the Aurora-2 connected digit database. Experimental results show that these NMF-based methods can improve the quality of noise-corrupted speech signals, and they are well additive to two well-known methods, spectral subtraction (SS) and minimum mean-squared error (MMSE). We also show that the speech signals enhanced by NMF-based methods can result in better recognition accuracy relative to the original signals without enhancement.[[note]]碩