一個結合SVM與Eigen-MLLR新的多語者線上調適架構應用於泛在語音辨識系統

A New On-Line Multi-Speaker Adaptation Architecture Combining SVM with Eigen-MLLR for Ubiquitous Speech Recognition System

本論文提出了一個以結合SVM和Eigen-MLLR為基礎的線上多語者調適架構，應用於ubiquitous環境的語音辨識系統。語者獨立式的辨識系統相較於傳統的辨識系統有著更好的辨識效果，而語者調適方法便是其關鍵所在。本論文應用SVM和Eigen-MLLR的特性作為調適技術的基礎，對於每個訓練語者的個別訓練語料做分類以及建立特徵參數向量空間。在語音辨識時，使用SVM找出測試語者所屬的類別，再找出類別相對應的MLLR特徵參數矩陣，並將其與非語者獨立模型結合成語者獨立模型。最後再利用辨識結果與原本的MLLR matrix和Eigenspace採取比重運算，並將運算結果更新原本的MLLR matrix。相較於非語者獨立的辨識系統可以增加了5~8%的辨識率。

This work presents a novel architecture using SVM and Eigen-MLLR for rapid on-line multi-speaker adaptation in ubiquitous speech recognition. The recognition performance in speaker independent system is better than in conventional speaker dependence system, and the key point is speaker adaptation techniques. The adaptation approach is on the basis of combine SVM and Eigen-MLLR, generating a classification model and building parameters vector-space for all speakers' individual training data. While in recognition, to find test speaker classification by SVM and look for MLLR parameters matrix correspond to speaker classification, then the MLLR parameters matrix and original acoustic model will integrate into speaker dependent model. Last, we estimate the adapted MLLR transformation matrix set by weighting function with recognition result, the present MLLR matrix, and Eigenspace. The estimate result will be used to update the MLLR matrices in adaptation phase. The experimental results show that the proposed method can improve 5% to 8% speech recognition accuracy with speaker adaptation.