Analysis on individual differences in automatic transcription of spontaneous presentations.pdfVIP

ANALYSIS ON INDIVIDUAL DIFFERENCES IN AUTOMATIC TRANSCRIPTION OF SPONTANEOUS PRESENTATIONS Takahiro Shinozaki and Sadaoki Furui Tokyo Institute of Technology Department of Computer Science 2-12-1 Ookayama, Meguro-ku, Tokyo, 152-8552 Japan. {staka, furui}@furui.cs.titech.ac.jp ABSTRACT 2. SPEECH RECOGNITION TASK AND EXPERIMENTAL SET UP This paper reports an analysis of individual differences in spon- 2.1. Recognition task taneous presentation speech recognition performances. Ten min- For the analysis of speaker variation, monologue presentation utes from each presentation given by 50 male speakers, for a total speech uttered by 50 different male speakers is used as a test set. of 500 minutes, has been automatically recognized for the analy- Speakers in the test set have no overlap with those in the training sis. Correlation and regression analyses were applied to the word set. The first 10 minutes of each presentation are used for analysis. recognition accuracy and various speaker attributes. A restricted Table 1 shows the detail of the test set. set of the speaker attributes comprising the speaking rate, the out of vocabulary rate and the repair rate was found to be most signif- 2.2. Speaker attributes icant to yield individual differences in the word accuracy. Unsu- We give consideration to seven kinds of speaker attributes in the pervised MLLR speaker adaptation worked well for im