Prototype LVQ Based Computerized Tool for Accent Diagnosis among Chinese Speakers of English as a Foreign Language

電腦化學習向量量化網路華人英語口音診斷系統之原型

杜凱蕾(Carol A. Troy) ； 傅家啟(Ja-Chih Fu) ； 黃經貿(Ching-Mao Huang)

《大葉學報》 8卷2期 (1999/12) Pp. 53-62

https://doi.org/10.7119/JDYU.199912.0053

摘要

母語為非英語系之地區，英語教學上現階段尚缺乏有效之設備或器材以輔助或矯正學生之英語發音。以音素為基礎之語音辨識提供一潛在可行之技術，以電腦輔助診斷成人常見之英語發音問題。本論針對華人英語發音，發展一以音素為基礎之語音處理系統（Chinese speakers English pronunciation tutorial, CEPOT）嘗試電腦化輔助診斷英語口音。CEPOT以音素為基礎之學習向量量化網路（learning vector quantization, LVQ）將以單字為單位之輸入訊號拆解為音素之組合，經與資料庫之標準發音比對後，CEPOT可偵測輸入單字之發音錯誤（如遺漏應有之音素或夾雜不應有之音素）並量測輸入發音各音素之清晰程度。實驗結果顯示CEPOT可有效偵測指定輸入單字中遺漏之應有音素或夾雜不應有之音素，但對發音清晰度之量測效果並不顯著。主要原因為CEPOT採用一般傳統所使用之傅立葉變換（Fourier transform）作為學習向量量化網路之前處理器，而傅立葉變換並非一有效工具以萃取語音等非穩態訊號。為改進上述演算法之限制，後續研究將致力於了解各音素發音與數位訊號間之相關關係，並期望進一部發展有效之演算法以量測英語發音之清晰度。

關鍵字

語音辨識 ； 學習向量量化網路 ； 電腦輔助語言學習

並列摘要

In Taiwan, university-level English as a Foreign Language (EFL) classrooms are usually an unsatisfactory venue for remedial pronunciation work. Recent advances in phoneme-based speech recognition have raised the prospect that Computer-Assisted Language Learning technologies could help fill this gap by giving adult EFL learners access to model pronunciations of problematic words and expert feedback on their own pronunciations. This paper presents CEPOT, a phoneme-based word recognition system that targets the pronunciation problems peculiar to Chinese EFL speakers. The core of CEPOT is an LVQ-based phoneme detector that enables it to derive the phonetic spelling of input words. By comparing the phonetic spellings of model pronunciations and student input pronunciations, CEPOT detects omission of necessary phonemes and addition of extra phonemes. By using a clarity index, CEPOT identifies unclear phonemes within the student pronunciations. Experimental results showed that CEPOT could successfully detect omission/addition of phonemes in selected whole-word inputs. However, it was not successful in measuring pronunciation clarity because its FFT-based preprocessor could not capture the non-stationary features of the speech signals. To improve CEPOT's performance, it will be necessary to better understand the relationship between the phoneme input, the signal characteristics, and the underlying speech production mechanisms.

並列關鍵字

speech recognition ； learning vector quantization ； computer aided language learning