臺灣閩南語元音間/t/變異音之聲學及構音研究

Abstract

許多語言蘊含有元音間輔音之交替現象。在臺灣閩南語的元音間/t/之處理過程中也會出現元音間輔音的交替現象。在詞內，元音間的/t/可能會保持原本的阻音[t]，也有可能會變成閃音[ɾ]，這種變化可能由於構詞規則的差異所致。本研究透過聲學與構音分析，探討相同構詞結構是否會引起相似的閃音化現象。本研究比較分析了三個元音間/t/的類型:詞內阻音(A類)、詞內閃音(B類)以及詞界間閃音(C類)。 本研究招募了八位臺灣閩南語使用者進行實驗。實驗過程中，蒐集三個不同類型的發音及超音波舌位影像，並分析其聲學及構音上的特徵。聲學分析結果顯示，阻音之形態呈現典型的[t]或其同位異音[ʔ]，而閃音的形態則呈現[ɾ]的特徵。在三個類型中，閃音的形態具有相似的時長及強度，而阻音的聲學特徵較為多元。另外，考量閃音較短的時長，本研究採用超音波M-Mode進行構音分析，並使用CoAtNet-2提取超音波影像之特徵向量以進行超音波舌位影像的量化分析，進而比較其餘弦相似度。該分析結果顯示[t]及其同位異音各有不同構音特徵，形態相似語境下的閃音(B類及C類)之間相似度高於不同環境下的[t]或[ʔ]的相似度。 這些發現證實了閃音化現象可能受到構詞條件的影響。研究結果也進一步提供了電腦視覺模型在分析超音波舌位影像時的實用性。

Intervocalic consonant alternation is a characteristic feature of many languages, and Taiwan Southern Min demonstrates this phenomenon, particularly in its treatment of /t/. Within words, the intervocalic /t/ either maintains its original posture [t] or becomes a flap [ɾ], potentially due to morphological differences. This study investigates whether morphological structures trigger comparable flapping using acoustic and articulatory analyses. Three structures were analysed: non-flapping intervocalic /t/ within words (category A), flapping within words (category B), and flapping at word boundaries (category C). Eight Taiwan Southern Min speakers were recruited for a production experiment, where they pronounced words and phrases from each category. Data collection included both acoustics and ultrasound tongue images. Acoustic analysis results showed that morphological structures forbidding flaps displayed the typical plosive [t] or its allophone [ʔ], while those allowing flaps exhibited characteristic [ɾ]. Consistent duration and intensity were observed for flaps whereas more variations were reported for the non-flapping category. On the other hand, ultrasound M-mode imaging was employed for articulatory analysis considering the brief nature of flaps. CoAtNet-2's pre-trained model was utilized to extract feature vectors and compare cosine similarity to address quantification challenges in ultrasound tongue imaging. Results showed characteristic patterns for [t] and its allophones, with flaps in morphologically identical contexts (categories B and C) showing higher similarity to each other than to [t] or [ʔ] in different contexts. These findings confirm that articulatory gestures of flapping may be morphologically conditioned. The results also highlight the utility of deep learning-based computer vision models for objectively analyzing ultrasound tongue imaging.