Performance Evaluation of Speaker-Identification Systems for Singing Voice Data

Computational Linguistics and Chinese Language Processing Vol. 16, No. 1-2, March/June 2011, pp. 1-14 1 © The Association for Computational Linguistics and Chinese Language Processing Performance Evaluation of Speaker-Identification Systems for Singing Voice Data Wei-Ho Tsai∗ and Hsin-Chieh Lee∗ Abstract Automatic speaker-identification (SID) has long been an important research topic. It is aimed at identifying who among a set of enrolled persons spoke a given utterance. This study extends the conventional SID problem to examining if an SID system trained using speech data can identify the singing voices of the enrolled persons. Our experiment found that a standard SID system fails to identify most singing data, due to the significant differences between singing and speaking for a majority of people. In order for an SID system to handle both speech and singing data, we examine the feasibility of using model-adaptation strategy to enhance the generalization of a standard SID. Our experiments show that a majority of the singing clips can be correctly identified after adapting speech-derived voice models with some singing data. Keywords: Model Adaptation, Singing, Speaker Identification. 1. Introduction As an independent capability in biometric applications or as part of speech-recognition systems, automatic speaker-identification (SID) (Rosenberg, 1976; Reynolds & Rose, 1995; Reynolds, 1995; Campbell, 1997; Reynolds et al., 2000; Bimbot et al., 2004; Nakagawa et al., 2004, 2006; Murty & Yegnanarayana, 2006; Matusi & Tanabe, 2006; Beigi, 2011) has been an attractive research topic for more than three decades. It is aimed at identifying who among a set of enrolled persons spoke a given utterance. Currently, existing SID systems operate in two phases, training and testing, where the former models each person's voice characteristics using his/her spoken data and the latter determines unknown speech utterances based on some comparisons between models and utterances. As the purpose of SID is distinguishing one ∗Department of Electronic Engineering & Graduate Institute of Computer and Communication Engineering, National Taipei University of Technology, Taipei, Taiwan Tel: +886-2-27712171 ext. 2257 Fax: +886-2-27317120 E-mail: whtsai@ntut.edu.tw The author for correspondence is Wei-Ho Tsai. 2 Wei-Ho Tsai and Hsin-Chieh Lee person's voice from another's, it is worth investigating if an SID system can not only identify speech voices but also singing voices. There are a number of real applications where an SID system may need to deal with singing voices. For example, if we record the sounds from TV, it is very likely that the recording contains performers speaking then singing or singing then speaking. In such a case, an SID system capable of handling both speech and singing voices would be very useful to index the recording. Another example is when people gather to sing at a Karaoke. It would be helpful to record everyone’s performance onto CDs or DVDs to capture memories of the pleasant time. For the audio in CDs or DVDs to be searchable, audio data would preferably be written in separate tracks, each labeled with the respective person. In this case, an SID system capable of identifying both speech and singing voices will …

Content truncated to 500 words