The Construction and Testing of a Mandarin Emotional Speech Database

The Construction and Testing of a Mandarin Emotional Speech Database Tsang-Long Pao, Yu-Te Chen, Jhih-Jheng Lu, Jun-Heng Yeh Department of Computer Science and Engineering, Tatung University, Taipei E-mail: tlpao@ttu.edu.tw, d890600S@mail.ttu.edu.tw, g9106001@ms2.ttu.edu.tw Abstract. Researches in speech synthesis and speech analysis are underpinned by the databases they used. The performance of an emotion classifier relies heavily on the qualities of the training and testing data. A good database can make researches in these fields achieving better results. Hearing-impaired people are poor in presenting their emotions in speech. We want develop a computer-assisted speech training system that can help to teach them to present their emotions similar to normal people. In this paper, we present a way to build a Mandarin emotional database, including the process of collecting data, arranging data, clips naming rules, and a listening test. Then we construct a computer-assisted speech training system to help in teaching the hearing-impaired people presenting their emotion in their speech correctly by analyzing the emotion in their speech and those in the database using KNN and M-KNN techniques. Keywords: Emotional speech database, Emotion evaluation, Emotion Radar chart, M-KNN 1. Introduction The performance of an emotion classifier relies heavily on the quality of emotional speech data and the similarity of it to real world samples. As mentioned in [1], there are three different categories of emotional speech: acted speech, elicited speech, and spontaneous speech. In this section we will describe the ways to obtain these three kinds of speech data. In acted speech recording, actors are invited to record utterances, where each utterance needs to be spoken with multiple emotions. The method is adopted by most researches because it can get large amount of data in a short time and the data is undistorted. For general use, we should invite speakers with different age, gender, even with different social or educational background if possible. And if we hope the emotion in the data to be more obvious, we could invite professional actors. We can also collect the clips that contain utterances with specific emotion in a film. We must avoid the background noise including music, surrounding noise, and other people’s voice. This method takes quite a lot of time in viewing the content of films. In elicited speech recording, the Wizard-of-Oz (WOZ) is used. The WOZ means using a program that interacts with the speaker and drives him into a specific emotion situation and then records his voice. This method needs a good program that can induce the participator to say something in our expected emotion state. So how to design such a program may not be easy. In spontaneous speech recording, the real-world utterances that express emotions are recorded. Although data got from this method has the best naturalness, it is the most difficult because we need to follow the speaker. When he or she is in some emotion state, his voice is recorded immediately. This method will face many problems. For examples, we must hide our recording device in order to make the …

Content truncated to 500 words