重音現象對於重音語言之理解語意判斷是一項重要的指標。為了發展德語電腦輔助語言學習系統中的輕重音判斷模組,本論文考慮韻律參數會受語句內容影響的特性,提出了一套正規化處理流程與參數求取方法。主要為基於Fujisaki Model分析基頻的分解能力,作為移除語句內容的影響,結合傳統正規化方式。並且以超音段(suprasegment)為單位,考慮目標母音與前後母音的相對關係以擷取特徵參數。 實驗採用「The Kiel Corpus of Read Speech, Vol. I」語料庫進行測試,利用二元樹篩選有效參數以求取最佳參數組合,進一步結合韻律和頻譜參數相互比較其辨識率。實驗結果顯示,相較於傳統的韻律參數擷取,本論文提出的正規化程序及特徵參數擷取方法,能夠有效的降低語句內容的影響。
Stress phenomenon is an important issue for the understanding of the stress-timed language semantic. For developing the stressed/unstressed judgement module of the german computer assisted language learning system, and considering the characteristics that prosody feature varies with the sentence content. A new normalization procedure and feature extraction method is proposed in this paper. Mainly based on the ability of fundamental frequency decomposition of Fujisaki Model, as remove the phrase influence. Moreover, extract features by considering the difference between the target syllable and it’s neighbors. The performance of the method is evaluated using 「The Kiel Corpus of Read Speech, Vol. I」database. Using decision tree for feature selection. Comparing to traditional feature extraction, the proposed methods is better and promising to reduce the phrase influence.