本研究計畫主要的工作為語者辨識，先進行快速的非特定文字語者辨認系統，在辨識模型參數的數量上有嚴格的限制，因此本計畫針對這些限制提出有效的解決方案。在語音特徵方面萃取語音訊號的小波轉換參數；在辨識器方面選用多維矩形類神經模糊系統。多維矩形類神經模糊系統(Hyper Rectangular Neuro-Fuzzy System, HRNFS)具有複雜決策邊界的辨識器，藉由誤差的傳遞修改權重值，並採用監督式決定導向學習 (Supervised Decision-Directed Learning, SDDL) 演算法訓練多維矩形類神經模糊系統，很容易以If-then的形式萃取分類規則。測試語料以德州儀器與麻省理工學院(TIMIT)共同開發的語音資料庫為主，實驗結果顯示本系統有很好的辨識效果。 The objective of this project is to find an effective middleware of speaker identification. It is to search effective methods of text-independent speaker identification with the constraints on the limited parameters of the model. We extract a set of features of each speech frame from wavelet transform and use hyper rectangular neuro-fuzzy system (HRNFS) as the classifier. HRNFS is a classifier with complicated decision boundary. The weighted values were tuned by propagation errors, and the HRNFS was trained by supervised decision-directed learning (SDDL) algorithm. It is easy to represent rules by If-then. In our experiments, the speech database is the Texas Instrument / Massachusetts Institute of Technology (TIMIT) acoustic-phonetic corpus of read speech. The effectiveness of the proposed system is confirmed by the experimental results.
