The Study on Emotion Recognition in Continuous Mandarin Chinese Speech

• Main Authors: Jun-Heng Yeh, 葉俊亨
• Other Authors: Tsang-Long Pao
• Format: Others
• Language: en_US
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/8fq8rx

博士 === 大同大學 === 資訊工程學系(所) === 98 === In this dissertation, 14 emotions, including 6 full-brown emotions: anger, boredom, fear, happiness, neutrality, and sadness, 8 underlying emotions: anxiety, despair, disgust, gratefulness, love, praise, shame, and surprise, are explored for speech emotion recognition study using speech processing and pattern recognition methods in the emotional space theory. In this study, a new emotional corpus, Mandarin Chinese Emotional Corpus 2010 (MCEC2010), is built up by 34 professional voice actors/actresses. This corpus consists of 19,136 short utterances recorded by voice actors/actresses and 472,017 emotional annotations labeled by 26 annotators who didn’t participate in the voice collection. Additionally, 209 dialogue scripts were recorded that were expressed freely by voice actors/actresses and annotated by 21 annotators who didn’t take part in recording these scripts. In the past, most of speech emotion recognition researches focused on recognizing emotions based on a single utterance as the recognition unit. This kind of sentence-type emotion recognition method may be inadequate to work in practice without the association of speech recognition mechanism. Moreover, the recognition rate will be affected by the accuracy of speech recognition in a hierarchy recognition scheme. Additionally, it must be noted that the heavy computation required for hierarchy recognition scheme would make it difficult to implement the emotion recognition in real time in a device with low computing ability. To overcome this shortcoming, four different speech segmentation methods for emotion recognition are tested in this dissertation. It appears that few research findings are available concerning to the analysis of mix-emotion in an utterance on emotion recognition. Most of them can only recognize the main emotional expression of the utterance. In this dissertation, the relative perceiving intensity of different emotional expressions of the segment is analyzed. And all emotional expression of segments will be mapped into the PAD (Pleasure-Arousal-Dominance) emotional space. By this visualizing presentation, it can not only quantify different emotional expressions in the utterance, but observe the gradual changes of different emotional expressions during dialogue.