結合聲調辨認之中文關鍵詞辨認系統

Abstract

在現今的國語語音辨認系統中，大多使用國語411音作為辨認單元，使用聲學模型後，大部分同音不同聲調的情況都能辨認出正確結果。但是在關鍵詞辨認系統中，多數情況關鍵詞都是命名實體(Named entity)，如人名、地名、公司名，常常都是二字詞且容易有混淆音，所以加上聲調辨認就十分重要了。 本論文中，使用兩階段式的keyword spotting系統，在原來關鍵詞辨認系統中之語音參數抽取部份，使用RAPT(A Robust Algorithm for Pitch Tracking) 演算法[1]，求取基頻軌跡，在系統辨認出Top-10 keyword後的likelihood分數，對關鍵詞加上第二級的MLP聲調辨認器[2]所辨認出來的分數，與Top-10加總後的分數，再進行重排，以得到更正確的辨認答案。 論文中，對一特定關鍵詞組：新竹科學園區341公司名（若包含別名則有1074個關鍵詞），製作關鍵詞辨認系統；在未加入聲調辨認時，關鍵詞辨認率為94.54％，加入第二級關鍵詞的聲調辨認器後，關鍵詞辨認率提昇至為95.32％，錯誤下降率為14.3％。

Most of today's Mandarin speech recognition systems use 411 syllables (regardless of tone information) as recognition unit, and most of them could be recognized correctly with the help of language model. However, in the case of keyword spotting, keywords are always Named Entities, such as person names, location names, company names,…, etc. Those keywords are usually only two characters in length and easily confused with each other. So it is important to recognize words with tone information. In this thesis, two-stage keyword spotting system is used. RAPT (A Robust Algorithm for Pitch Tracking) is applied to get the pitch contour in the feature extraction phase of the original system. The likelihood scores derived from Top-10 keyword recognition are added with the scores from the second stage MLP tone recognizer, and then the scores with Top-10 results are reordered to get better recognized answers. In this thesis, keyword spotting system is made for a specific keyword phrases: 341 company names (1074 including the aliases) in Hsinchu Science Park. The keyword recognition rate is 94.54% without tone recognition, which increases to 95.32% with the second stage tone recognizer, and the error reducing rate is 14.3%.