Influences of Prosodic Feature Replacement on the Perceived Singing Voice Identity

Influences of Prosodic Feature Replacement on the Perceived Singing Voice Identity Kuan-Yi Kang Department of Electrical Engineering National Tsing Hua University sjk102061231@gapp.nthu.edu.tw Yi-Wen Liu Department of Electrical Engineering National Tsing Hua University ywliu@ee.nthu.edu.tw Hsin-Min Wang Institute of Information Science Academia Sinica whm@iis.sinica.edu.tw Abstract Human perception on the singing voice differs with the factors of the singing voice and the subjects. On one hand, the background knowledge influences the understanding of voice for each subject. On the other hand, the difference of the voices presented to the subjects also affects the perception. In this paper, we discuss two factors reflecting on the similarity before and after singing voice conversion: prosodic features and subjects’ familiarity to the singers. Three experiments were conducted. The first experiment tested the subjects’ ability to identify the singer. The second experiment synthesized the singing voice with different singers’ prosodic features, and let the subjects score the similarity. The third experiment presented timbre-converted singing voice with different combinations of prosodic features from two singers to the subjects for them to judge the similarity to the target singer. The results show that, first, the number of prosodic features contained in the synthesized voice is positively correlated with the scores in identification and similarity. Also, subjects who are more familiar personally with the target singers have better identification scores than target-unfamiliar subjects on the timbre-converted singing voices. Keywords: Singing voice conversion, Prosody, Human perception, Voice identity. The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 296-309 ©The Association for Computational Linguistics and Chinese Language Processing 296 1. Introduction In the task of voice conversion, subjective tests examining quality and similarity [1] are used to evaluate the synthesized results from human perception. The quality test asks the subjects to score how great the quality of the converted result is, while the similarity test questions about how similar the result is, comparing to the target speaker. The acoustic characteristics of voice individuality can be described through timbre, pitch, intensity, duration, etc. [3]. Hence, the synthesis of the voice considers not only the timbre but also other prosodic features such as pitch, duration or intensity [4]. What effects the modification of these features have on the identification of speakers [5] and how these features represent individuality [6,7,8] have also been discussed in previous works. Studies in voice perception have investigated on the acoustic features affecting voice identity with experiments using the correlation analysis and multidimensional scaling techniques [12]. The modifications of acoustic features [13,16] and vocal identity aftereffects [14] are also used to determine the relative importance factors on humans’ ability of speaker identification. While some studies on the contributing features of voice identity contain timbre and other prosodic features [12,13,16], this paper investigates only on the prosodic features in order to find out the essential prosodic parameters for changing the perceived singer identity. In addition, since that the majority of the voice conversion tasks only focus on the development of the timbre conversion [1], using the source prosodic features …

Content truncated to 500 words