語音辨識字幕對口譯學生英進中同步口譯數字與專有名詞處理之影響

Abstract

本研究旨在探討學生口譯員英進中同步口譯時，線上免費語音辨識字幕對數字與專有名詞處理之影響，透過實驗比較有語音辨識字幕與沒有語音辨識字幕情況下，口譯學生表現之差異。實驗素材為兩篇長度九分半鐘的英文演講，以錄音檔的形式播放演講音檔，一篇音檔單獨搭配投影片播放，一篇音檔則於投影片旁播放語音辨識字幕。由於學生口譯員普遍較容易使用免付費的語音辨識服務，故選用免費的Otter.ai軟體產出英文語音辨識字幕，並在參與者口譯有語音辨識字幕的音檔時，以眼動軟體追蹤參與者目光。本實驗邀請13位臺灣的口譯碩士生參與研究，參與者完成同步口譯實驗後，進行刺激回想訪談，回想口譯時的挑戰與採用之策略，以此探究語音辨識字幕是否影響訊息處理以及口譯產出。 本實驗將參與者口譯內容進行分析，首先替譯文中的數字評分，發現在有語音辨識字幕的情況下，參與者獲得的平均分數較高，而且成對樣本t檢定結果顯示達顯著水準。研究結果也發現數字的錯誤率在有語音辨識字幕的情況下較低。本研究亦針對含數字與專有名詞的句子以及相鄰的句子進行準確度評分，結果顯示在有語音辨識字幕的情況下，句子準確度的平均分數較無字幕的平均分數高，然而成對樣本t檢定結果顯示此平均值無顯著差異。本實驗透過刺激回想訪談的結果，探究參與者對於口譯搭配語音辨識字幕應用之想法。結果顯示參與者對同步口譯搭配語音辨識字幕持保留態度，部分人認為語音辨識字幕在特定情況下有助於口譯表現。本研究最終參考氣力模型，探討字幕對同步口譯氣力分配之影響。

This empirical study investigated the effects of unedited automatic speech recognition (ASR) transcript on student interpreters’ simultaneous interpretation (SI) performance from English into Chinese in terms of numbers and proper names. The experiment recruited thirteen student interpreters who took part in an SI experiment and a following retrospective interview. The SI experiment required the participants to interpret two speeches, one with the presence of unedited ASR transcript and one without. During the tasks with ASR, a laptop displayed slides accompanied by ASR transcripts generated by Otter.ai, and an eye-tracking software was adopted to capture the participants’ eye movements during the task. In the tasks without the ASR, only slides were provided. Each of the two speeches in the SI experiments contained 20 numbers and 10 proper names. After the interpreting tasks and the retrospective interviews, the participants were asked to fill in a post-experiment survey and rate the perceived effectiveness of the ASR. Quantitative analysis was adopted to analyze the number renditions and the accuracy of sentences containing problem triggers and the neighboring sentences. A paired sample t-test showed that the participants scored significantly higher in number renditions when ASR was present. The accuracy of sentences in the tasks with ASR received higher scores, but a paired sample t-test showed there was no significance difference between the scores of the tasks with ASR and without ASR. In terms of proper names, results showed there were fewer omissions in the tasks with ASR. Finally, the retrospective interviews revealed that the participants adopted a more neutral stance towards the application of unedited ASR service, and they maintained that the ASR support might be useful under certain circumstances. Gile’s Effort Models were discussed at the end of this paper with an aim to explain the impact of unedited ASR on an SI task.