Abstract
本篇論文中將探討幾種語者識別的方法，它們分別是鑑別訓練法、新的計分方法以及音段機率模式。在鑑別訓練法中，我們對每一個語者均建立一個隱馬可夫模式，借由考慮其他競爭語者的效應，鑑別訓練法的提出，將可擴大語者間的差異性。為了達到鑑別訓練法的目的，廣用機率階梯方法將被用以估算新的語者模式。 為了更進一步改善傳統在語者識別上的計分方法，本文中提出了另一種計分方法，這種計分方法的原理是來自貝斯最小冒險檢測，其目的是為達到最小錯誤機率。 傳統的音框基準機率模式是使用每個音框的瞬間頻譜特性作為語者識別的特徵，為了提高識別率，我們提出一種新的段落機率模式，此種模式同時使用了瞬間頻譜及暫態頻譜訊息。固定長度的段落機率模式雖可提高語者識別率，但是它不能有效的代表語者的特性，本論文中，我們將更進一步提出一種聲段基準的模式，在此模式下，語音信號的特徵將用正交函數來代表，一種漸進的方法被用以切割及模式化語音信號，經由此一漸進方法，語音段落的邊界及語者模式均可被自動找出及求出。本論文中所使用的語料庫為100個人的語料，計有50個男及50個女，語料內容是英文的數字資料厙。在每一章節中，均有一些實驗去測試上述所提出的方法。This dissertation investigates the techniques in speakerrecognition. They are the discriminative training algorithm,the scoring function and the acoustic segment basedprobabilistic model.Some new methods are derived for improvingthe accuracy rate. In discriminative training,we construct thehidden Markov model for each speaker.By taking into accountthemodels of other competing speakers so that speaker separationis enhanced. The optimization solution can be obtained by usinga probabilistic descent algorithm. In order to improve theperformance of conventional scoring algorithm, we propose a newscoring method for speaker verification.This method is derivedfrom the Bayes test for minimum risk to attain the objective ofminimizing the error probability. The conventional frame basedprobabilistic models use the instantaneous spectral informationof each individual frame. Instead, we proposed a new segmentmodel which uses both instantaneous spectral information andprevious spectral information. The fixed length segment modelis a simple approach to get performance improvement. A moreprecise method is to consider the acoustic segments. In thismethod, speech signals are represented by the orthogonalpolynomial function. An iterative algorithm is proposed tosegment and model the speech signals. The segment boundariescan be automatically detected according to the characteristicsof speech signals so that we can generate the segment basedprobabilistic models. A 100-speaker digit database collectedthrough the public switching telephone network is used for aseries of experiments to show the effectiveness of our proposedmethods.