反傳遞網路應用於語音及語者辨識

Abstract

語音及語者辨識已經被發展了許多年，也已有不少的論述發表，而較常用 的辨識方法大多是屬於動態時間校準和隱藏式馬可夫模型這兩種。近年來 ，由於類神經網路的幾個特點：(i) 平行處理的能力，(ii) 學習的功能 ，(iii) 容忍錯誤；使得其越來越受到注意並被應用於語音於的處理，反 傳遞網路即是類神經網路的一種。本論文提出採用一個反傳遞網路的變形 －前向反傳遞網路來應用於語音及語者辨識以作為一個新的嘗試。在本論 文裡分別將此種前向反傳遞網路應用在語音及語者的辨識之上，最後再將 此二者合併成一個”語者－語音”辨識系統以期提高辨識率。由於獲得了 令人滿意的結果－高辨識率及快速的辨識時間，使我們相信是值得去花更 多的努力在應用反傳遞網路於語音及語者辨識之研究上。 Speech and Speaker recognition are studied for many years, and there are have been many researches on that. Most of those methods on speech are developed from DTW (Dynamic Time Warping) or HMM (Hidden Markov Model), and those researches are done well enough and hard to get further more development. In this years, their are more and more people begin to use ANN (Artificial Neural Network) in speech processing because of ANN' s several characteristics: (i) parallel processing, (ii) ability of learning, (iii) fault tolerance. The CPN (CounterPropagation Network) is a kind of ANN's model presented by Robert Hecht- Nielson on 1987. This thesis presents using a variant of the CPN model, the farward-only CPN, in the applications of speech and speaker recognition for a new attempt, and there are no other study for using CPN in speech and speaker recognition yet. In this thesis, separately apply the forward-only CPN in the speech recognition and speaker recognition, and then combine these two system into one speaker- speech recognition system that can satisfy to the requirement for real-time. The perfoarmance exceeds our expectation that it get a high accuracy and a short recognition time, so it is worth to pay more efforts in the researches for using the CPN in speech and Speaker's recognition.