Automatic personality perception from speech in Mandarin

Researches on speech-based personality perception have been actively conducted in recent years. However, most of them are focus on French, while research on Chinese corpus has not been reported. This paper investigates automatic perception of speakers' personality from speech in Mandarin. The studied personality traits are extended from standard Big-Five to 6 sub-traits of each, leading to a total of 30 traits. As for modeling approach, traditional SVM system with standard prosodic feature extraction is studied as baseline. A novel skip-frame LSTM system is proposed, which resorts to skip-frame sampling to augment training data while retaining prosodic variation across a long time. The LSTM system learns personality information directly from frame level descriptor (MFCCs) instead of manually designed prosodic features as SVM system does. The results of both the two systems show that, the extraversion trait is the most easily perceived while openness is the most difficult one, which is in agreement with the results in French. In contrast, the high accuracy on agreeableness is peculiar to the previous result in French, which may relate to different culture backgrounds. In addition, relation between Big-Five and their sub-traits are discussed. Finally, the proposed skip-frame LSTM system outperforms obviously than SVM overall, demonstrating that LSTM can serve as an effective method for automatic personality perception even in low-resource condition.