廳堂之中文語音清晰度與訊號自函數之關聯性研究——以臺灣地區華語為主

A Study of the Chinese Speech Intelligibility in Halls in Relation to the Autocorrelation Function Model: A Case of Chinese in Taiwan

台灣目前測定廳堂語音清晰度品質之優劣，多以英語發音為基礎之快速語音傳輸指標(Rapid speech transmission index, RASTI)來作為聲場評價依據。陳炯堯等(2003)已經證實依據台灣地區的華語語音所構成之清晰度檢測音表的檢測(articulation test)與RASTI之量測值就大體而言是一致的；但就不同單音節而言發現少數發音個體具獨特差異性。因此室內廳堂語音清晰度的研究中，所考慮的因素不僅該包含空間組成的特性，另一方面也該慮單音節字音的發音特性。本研究目的為綜合上述兩種問題提出新的語音分析方法來解析RASTI不足之處，我們利用語音訊號的自函數(autocorrelation function, ACF)運算，來探討中文語音物理特性與檢測音表實驗結果是否有特殊的關聯性。在分析研究中利用7個聲場的空間特性，發現在84個觀測點比較分析中，就同一聲場而言，自函數的最小有效遲延時間(τe)min 與各測點間主觀心理實驗結果有良好的對應關係(p<0.01)。此外，進一步就單一聲場分析發現(τe)min 可有效檢測不同單音節的構造因發音時間部位的不同與自成音節之韻母發音有明顯的差異性(r =0.81, p<0.01)。

Today, the English-based rapid speech transmission index (RASTI) is a command tool used in Taiwan to serve as the basis for evaluating sound field in the measurement of speech intelligibility in halls. In the previous study (Chen et al., 2003), the results of Chinese-based articulation test of mono-syllabic test lists in halls were similar to those of the RASTI. However, there is a significant finding that among different monosyllables, there were few monosyllables possessed individual difference. Thus, in a study of speech intelligibility in halls, it is important to consider the quality of sound reception of halls as well as the pronunciation features of monosyllables. The aim of this study is to consolidate the two aforementioned questions to analyze the inadequacy of RASTI method and to discuss whether the physical properties of Chinese speech is correlated to results of test lists by proposing new speech analysis method such as autocorrelation function, ACF. In this study, there were 7 sound fields being surveyed, and by comparatively analyzing 84 survey points, the results indicated that, as far as a sound field is concerned, the minimum autocorrelation function (τe)min was significantly correlated (p<0.01) with the articulation scores collected from each survey point. Meanwhile, after analyzing the sound field further, the finding indicated that the minimum autocorrelation function (τe)min can effectively detect different construction of monosyllables that may vary significantly (r=0.81, p<0.01) due to different place of articulation and different vowel articulation.