以自組織特徵映射建立國語聲調電腦評量模式之研究

Research of Computer Evaluation of Mandarin Tone Based on SOM

陳桂霞(Kuei-Shya Chen) ； 黃重光(Chang-Kang Huang)

《測驗統計年刊》 9期 (2001/12) Pp. 169-192

https://doi.org/10.6773/JRMS.200112.0169

摘要

語言是上天賦予人類最自然的溝通方式，也是人與人之間最快速且便利的溝通工具。學會使用國語來表達自己的情意，更是每個國民應具備的基本條件。國語是一種聲調語言，也就是聲調在國語中具有重要的辨義作用。所以聲調對於要學好國語來說是非常重要的。任何的教學最後一定要有評量，評量一方面可以提供學習的人適當的回饋，另一方面可以供教學者當成下一次教學的參考。國語聲調的評量一直都是以人工評量為主，本文則嘗試發展一套以電腦來評量學生唸讀單音國字聲調的評量模式。本研究針對國小三年級的學生，唸讀二十四個國字音，經人工評量及電腦評量學生的聲調後，再比較兩種評量結果的一致性。研究者在既有的相關理論上成功的應用語音處理技術來建立聲調的電腦評量模式，本論文主要的成果如下：在國音學聲調方面，使用「國語單字音聲調評鑑表」，做為人工評量的量表，並且肯定了人工評量的可靠性及一致性。在擷取聲調的演算法上，研究者以自相關函數法為基礎，考慮開放空間的情形下，成功的將動態的起終點偵測(endpoint detection)及聲調平滑演算法應用在聲調的評量上。評量模式的建立上，以自組織特徵映射類神經網路模式加上聲調特徵建立起單一模式四個聲調的電腦評量，聲調評量回饋模式則為本評量模式的另一特色。而在電腦評量與人工評量的一致性上，第一聲約為0.92、第二聲約為0.86、第三聲約為0.82、第四聲約為0.95，整體的一效性則可達到0.89。

關鍵字

聲調 ； 自組織特徵映射 ； 電腦評量

並列摘要

Language is a unique way of communication in humans. It is also an effective and convenient means of communication between people. To utilize language, for example Chinese, to expressive our inner thoughts is the responsibility of our people. Chinese is tonal languages which states the idea that tone in Chinese convey semantic meanings. Therefore, it is important to learn to use the tone in our language precisely. Any types of teaching must involve evaluation. Evaluation provides learners a sufficient amount of feedback. On the other hand, it provides teachers a concrete criterion of setting teaching goals for the next teaching session. Chinese tonal evaluation used to be done by humans. This research developed a computerized evaluation to evaluate learners' tone in Chinese. Subjects of this research were the third grade in local elementary school. Each student read 24 Chinese characters. This processing each was evaluated by the traditional way and the computerized evaluation. The researcher for their consistency compared both evaluations. This speech signal process was used to establish computerized evaluation model in evaluating tones in Chinese and was based on the related field of studies. The results of this research was as follows: In Mandarin phonetics tones, the ”Evaluation sheet of Mandarin tones” was used in the human's evaluation. The result confirmed the reliability and consistency of human's evaluation. Researchers use the method of autocorrelation for extracting the features of tones based on the autocorrelation. In an open space, successfully applied endpoint detection and tone smooth computerization in tone evaluation. In establishing evaluation model, a signal computerized mode for each tones was created by using self-organizing feature maps. Tone evaluation feedback model is another characteristic of this research. In the consistency of computerized evaluation and human evaluation, first tone has the point of 0.92, second tone was 0.86, third tone was 0.82. the fourth tone was 0.95. The overall consistency rate was 0.89.

並列關鍵字

tone ； self-organizing feature map ； computerized evaluation