調變頻譜分解之改良於強健性語音辨識

Abstract

自動語音辨識(Automatic Speech Recognition, ASR)系統常因環境變異而導致效能嚴重地受影響；所以長久以來語音強健(Robustness)技術的發展是一個極為重要且熱門的研究領域。本論文旨在探究語音強健性技術，希望能透過有效的語音特徵調變頻譜處理來求取較具強健性的語音特徵。為此，我們使用非負矩陣分解(Nonnegative Matrix Factorization, NMF)以及一些改進方法來分解調變頻譜強度成分，以獲得較具強健性的語音特徵。本論文有下列幾項貢獻。首先，結合稀疏性的概念，期望能夠求取到具調變頻譜局部性的資訊以及重疊較少的NMF基底向量表示。其次，基於局部不變性的概念，希望發音內容相似的語句之調變頻譜強度成分，在NMF空間有越相近的向量表示以維持語句間的關連程度。再者，在測試階段經由正規化NMF之編碼向量，更進一步提升語音特徵之強健性。最後，我們也結合上述NMF的改進方法。本論文的所有實驗皆於國際通用的Aurora-2連續數字資料庫進行；實驗結果顯示相較於僅使用梅爾倒頻譜特徵之基礎實驗，我們所提出的改進方法皆能顯著地降低語音辨識錯誤率。此外，也嘗試將我們所提出的改進方法與一些知名的特徵強健技術做比較和結合，以驗證這些改進方法之實用性。實驗平台使用HTK與KALDI兩種語音辨識系統。前者用來實驗上述所提出NMF改良之效能；後者用來實驗類神經網路(Neural Network)技術於語音辨識之聲學模型的效能，並探討調變頻譜正規化法與其結合之效果。The performance of an automatic speech recognition (ASR) system is often severely deteriorated due to the interference from varying environmental noise. As such, the development of effective and efficient robustness techniques has long been a challenging research subject in the ASR community. In this thesis, we attempt to obtain noise-robust speech features through modulation spectrum processing of the original speech features. To this end, we explore the use of nonnegative matrix factorization (NMF) and its extensions on the magnitude modulation spectra of speech features so as to distill the most important and noise-resistant information cues that can benefit the ASR performance. The main contributions include three aspects: 1) we leverage the notion of sparseness to obtain more localized and parts-based representations of the magnitude modulation spectra with fewer basis vectors; 2) the prior knowledge of the similarities among training utterances is taken into account as an additional constraint during the NMF derivation; and 3) the resulting encoding vectors of NMF are further normalized so as to further enhance their robustness of representation. A series of experiments conducted on the Aurora-2 benchmark task demonstrate that our methods can deliver remarkable improvements over the baseline NMF method and achieve performance on par with or better than several widely-used robustness methods.