Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/7958
標題: 應用階層性遞迴糢糊類神經網路於噪音環境下的語音辨識
Noisy Speech Recognition by Hierarchical Recurrent Neural Fuzzy Networks
作者: 邱祺添
Chiu, Chi-Tien
關鍵字: Hierarchical;階層性;neural fuzzy network;Noisy Speech Recognition;糢糊類神經網路;噪音環境下的語音辨識
出版社: 電機工程學系
摘要: 
本論文將提出一個具有階層性遞迴模糊類神經網路(HRNFNs)用於噪音環境下的語音辨識。HRNFNs 其架構包含兩個遞迴模糊類神經網路(RNFNs),前半部分是運用 RNFNs來實現具有過濾語音雜訊的濾波器;後半部分則是運用RNFNs來完成辨識的動作。RNFN是個由一系列的若...則...(IF…THEN…)法則所發展的遞迴模糊類神經網路,其遞迴的屬性是他對於具有時序性的樣本能夠充分表達其特性。假設我們有n 組語音就會有n 組RNFNs,在RNFNs的訓練上我們針對每一種語音就訓練一個網路,在訓練中我們運用預估(prediction) 的方法來預測。由此刻的語音frame feature 預估下一個語音frame feature這樣反覆的預測直到所有的語音資料都訓練完畢,然後以預估值的誤差總合來做辨識的依據。而在RNFN濾波器的訓練也是相同,有 組語音就會有 組的RNFNs。在HRNFN架構方面每一種語音具有其一RNFN辨識器,和一個RNFN濾波器。HRNFNs 其運作是當語音在進入RNFN辨識器前先用RNFN濾波器來過濾具有雜訊的語音。在我們的實驗中是用特定語者所錄的語音再加入不同的雜訊來做辨識,比較其他的網路如:多層感知機(MLP)、時間延遲類神經網路(TDNN)、隱藏式馬可夫模型(HMM),其結果是HRNFN具有很不錯的性能效果。

Noisy speech recognition by Hierarchical Recurrent Neural Fuzzy Networks (HRNFN) is proposed in this thesis. The proposed HRNFN is a hierarchical connection of two RNFNs, where one is used for noise filtering and the other for recognition. The RNFNs are constructed by recurrent fuzzy if-then rules, and the recurrent property makes them suitable for processing speech patterns with temporal characteristic. In n words recognition, RNFNs are created for n words modeling. Each RNFN receives the current frame feature and predicts the next one of its modeling word. The total prediction error of each RNFN is used as recognition criterion. In filtering, n RNFNs are created, and each RNFN recognizer is connected with a corresponding RNFN filter, which filters noisy speech patterns in the feature domain before feeding them to the recognizer. Experiments on speaker dependent Mandarin words recognition under different types of noises are performed. Other recognizers, including multi-layer perceptron (MLP), time-delay neural networks (TDNNs), and Hidden Markov Models (HMMs), are also experimented and compared. Admiring results of HRNFN in noisy speech recognition task are demonstrated from the experiments and comparisons.
URI: http://hdl.handle.net/11455/7958
Appears in Collections:電機工程學系所

Show full item record
 

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.