國語單音節及詞組混淆量測模型在關鍵詞辨認之應用

Abstract

本論文嘗試以別於以往關鍵詞系統需要大量語料訓練關鍵詞音節模型及填充音節模型的方式，針對導致音節辨認發生錯誤的音節混淆特性，依據辨認觀測機率差值分佈，予以模式化而為音節混淆量量測模型。並在單詞是由音節間獨立組合而成的假設下，採用音節混淆模型直接建立單詞混淆量模型。而本關鍵詞系統，是將連續語音經過TOP-N辨認器產生辨認音節及切割位置，再由關鍵詞搜尋器，從候選音節中串接出可能的候選關鍵詞及其所對應不受辭典樹限制辨認而得的最佳非關鍵詞。於是我們便引用之前所建立的音節模型分別建立所有混淆詞組的混淆量分佈模型，而在單詞錯誤率要求的條件下，依據其單詞混淆組合的模型求取出個別的加分補償量。以將與非關鍵詞較具有單詞混淆特性的候選關間擷取出而成為關鍵詞。如此便能因關鍵詞系統對單詞錯誤率的要求計算出補償量，而擷取出輸入之關鍵詞。

In this thesis, a new keyword spotting system using reward function to discriminate keywords from the background fillers was proposed. It first defined a confusion measure between Mandarin syllable-pair as the HMM recognition score difference of the correct and alternate hypothesis. And, the Gaussian pdf is used to model the confusion measure of syllable-pair. Under the assumption of independence between adjacent syllables, the confusion measure of word-pair becomes the summation of corresponding syllable-pairs. Finally, the confusing measure of word-pair was used to decide the reward function in the keyword spotting system. In the proposed system, the Top-N syllable lattice was first found by HMM syllable recognizer, and the keyword candidate can be found from syllable lattice associate with the most probable filler model. And, the reward function of each keyword candidate can be decided from confusion measure between keyword-filler word-pair and the desired keyword recognition rate.