分頻式調變頻譜分解於強健性語音辨識

Sub-band modulation spectrum factorization in robust speech recognition

在本篇論文中，我們使用了非負矩陣分解（nonnegative matrix factorization, NMF）技術來強化語音特徵調變頻譜、藉此提升自動語音辨識系統之雜訊強健性，其中，NMF法為語音之調變頻譜的強度求取一組基底向量，而我們藉由此組基底向量來擷取語音中重要的辨識成分，跟以往基於NMF之強健技術不同之處在於兩點：其一，我們利用了正交投影（orthogonal projection）的方式取代原先的迭代方式，使運算速度大幅增加。其二，我們採取分頻帶分解的方式取代原先全頻帶分解，藉此減少計算量。在Aurora-2之連續數字資料庫之辨識實驗顯示，上述的新方法相對於基礎實驗而言，能有效提升雜訊環境下語音辨識的精確度，可提供高達58%的相對錯誤改善率，而跟原NMF法相較，新方法運算複雜度明顯降低，而能維持原辨識精確度、部分甚至有提升的效果。

This paper proposes a novel scheme that enhance the modulation spectrum of speech features in noise speech recognition via non-negative matrix factorization (NMF). In the presented approach, we apply NMF to obtain a set of non-negative basis spectra vectors which derived from the clean speech to represent the important components for speech recognition. The difference compared to the conventional NMF-based scheme that leverages iterative search to update the full-band modulation spectra is two: first, we apply the orthogonal projection to update the low sub-band modulation spectra. Second, we process the low half-band of the modulation spectrum rather than the full-band. The presented new process improves the computation efficiency without the cost of degarded recognition performance. In the Aurora-2 database and task, the presented new NMF-based approach can achieve the average error reduction rate of over 58% relative to the baseline MFCC.