臺灣大學: 電機工程學研究所連豊力李運鋼Li, Yun-GangYun-GangLi2013-03-272018-07-062013-03-272018-07-062011http://ntur.lib.ntu.edu.tw//handle/246246/253960為了可以實現更有效率以及更具可信度的人機互動，人工情感模型應用在人機互動中是個關鍵的要素。除此之外，由於語音在我們日常生活中是一種不可或缺而且相當便利的溝通方式，因此若以語音作為人機互動中最主要的溝通方式，並且從語音中讀取其情緒成分，應用於人工情感模型中，可期待人機互動將更為順暢。因此，設計一個具有感知語音情緒功能，並且能憑藉其類似於人們情緒的人工情緒進而做出不同反應的機器人，是大有可為的。 根據情緒心理學，各種不同的情緒皆處於同一個情緒網狀系統中，各自皆有某種機率去影響其他的情緒；而在隨機模型中的隱藏式馬可夫模型，也是將各狀態定義於機率分佈空間，以彼此之間的條件機率來描述各狀態之間的轉移過程。因此，在本篇論文中，所使用的人工情感模型，是個根據於隱藏式馬可夫模型特性所建構出的情感模型。論文中，以此模型模擬在沒有外界刺激時，內部情緒自我調整的動態過程；以及模擬連續受到相同語音情緒辨識結果所造成之激刺時，其情緒轉移的動態過程。除此之外，藉由選定特定的模型參數，可以得到其相對應的情緒特性曲線，進而分別建立出具有樂觀與悲觀個性特質的模型。 然而，有鑑於目前對於各種情緒的定義還是很模糊不清，並且其理論也尚未建立完善，因此，在現實生活中語音情緒辨識正確率將會有所受限。若將此問題列入考量，錯誤的辨識結果必然將會對於先前設計的個性特質造成重大的影響。 為了維持所設計的個性特質之一致性，在此論文中，將個性模型受到所有辨識錯誤組合的影響全部列出考慮，並且經由一連串的模擬，分析並且探討出各種辨識錯誤組合與模型參數之間的關聯性。最後，將所有辨識錯誤組合分為數個組別以方便討論，再藉由調整每個組別其相對應的模型參數，將其對個性特質的影響，修正到容許的誤差範圍內。Artificial emotion model is considered as a key factor to achieve a more effective and believable human-robot interaction. As the speech communication plays an essential part of our daily life, the utilization of the emotional speech is expected to make human-robot communication smooth. Thus, it is promising to design a robot which has perception of emotion in speech, and responses corresponding to internal emotion similar to human. Since the individual emotion exists as a part of an emotion network and they have certain probability to interact with other emotions, HMM (Hidden Markov Model) as a stochastic model is an appropriate way to describe transition process of emotions. Therefore, an emotion model based on HMM is used in this thesis. The model simulates the dynamic processes of emotional self-regulation and emotional transference under the influence of the same stimuli arouse by the result of emotional speech recognition. In addition, by selecting the parameters of the model, the models of optimistic and pessimistic personality traits are also built. However, since the theory of emotion definition is obscure and not well-established, the accuracy of emotional speech recognition in real life is limited. Consequently, the error of recognition will have impact influence on the characteristic of designed personalities. In order to maintain the consistency of the personality traits, the personality models under the influence of all the combinations of recognition error are considered and analyzed to figure out the relationship with model parameters through a series of simulation. At last, they were divided into groups to discuss with for simplification, and each group is modified by adjusting the corresponding model parameters within acceptable error requirement.1763825 bytesapplication/pdfen-US語音情緒辨識情感模型隱藏馬克夫模型人機互動emotional speech recognitionemotion modelhidden markov modelhuman-robot interactionthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/253960/1/ntu-100-R97921050-1.pdf