基於多實例學習法之自然語音情緒辨識

Abstract

近年來因智慧型攜帶裝置逐漸普及，造就人們接觸智慧型人機介面的機會越來越多，但其情緒議題一直以來都被漠視，於是，擁有情緒的智慧型人機介面這課題就在人們快速膨大的需求中浮出。在心理學領域長期的研究下，關於情緒的研究已經有大量的資料可供參考。 近十幾年來由於情緒計算(Affective computing)的發展，在資訊工程領域方面也漸漸的累積了許多研究。以前人的研究做為基石，本研究選擇進行聲音方面的情緒辨識，為了簡化問題，本研究並不考慮說話內容，而單純就語氣和音調來辨識情緒。本研究採用個人簡報作為環境來收取自然聲音，並將情緒辨識集中在緊張情緒程度辨識上。在標記方面，本研究為了降低標記者的負擔，提出了以比較方式進行標記的解決方法。在辨識方面，本研究在各種的資料組合下比較一般傳統語音情緒辨識和MI-SVM的準確率。就結果方面，相較於傳統SVM辨識只有66%的準確率，本研究提出使用MI-SVM並獲得74%準確率。未來將以本研究為基石進而探討新的辨識方法的方向。

Because of the popularizing of smart carrying equipment, the chance of people to approach agents is increasing, and the issue of given agent emotion emerge from the enormous volume of increasing demand. In this research, we choose the the recognition of emotion in speech. For simplify the problem of recognition emotion, we ignore the content of speech, and simply recognizing emotion from the tone of speech. we choose the presentation environment as the research target to collect speech sound, and focus on the recognition of levels of nervous emotion. At the aspect of annotation, in order to reduce the burden of annotator, we propose a novel solution of annotation by comparing speech turns. At the aspect of recognition, we try a lot of combination of data type for comparing the recognition accuracy of traditional SVM method with the recognition accuracy of MI-SVM. As a result, the recognition accuracy of traditional SVM is only 66%, and the recognition accuracy of proposed MI-SVM is 74%. In the future, we will base on this research to find new way of recognition.