Committee-Based Active Learning for Speech Recognition

• HAMANAKA Yuzo

  Tokyo Institute of Technology

• SHINODA Koichi

  Tokyo Institute of Technology

• TSUTAOKA Takuya

  Tokyo Institute of Technology

• FURUI Sadaoki

  Tokyo Institute of Technology

• EMORI Tadashi

  NEC Corporation

• KOSHINAKA Takafumi

  NEC Corporation Tokyo Institute of Technology

Abstract

We propose a committee-based method of active learning for large vocabulary continuous speech recognition. Multiple recognizers are trained in this approach, and the recognition results obtained from these are used for selecting utterances. Those utterances whose recognition results differ the most among recognizers are selected and transcribed. Progressive alignment and voting entropy are used to measure the degree of disagreement among recognizers on the recognition result. Our method was evaluated by using 191-hour speech data in the Corpus of Spontaneous Japanese. It proved to be significantly better than random selection. It only required 63h of data to achieve a word accuracy of 74%, while standard training (i.e., random selection) required 103h of data. It also proved to be significantly better than conventional uncertainty sampling using word posterior probabilities.

Journal

• IEICE Transactions on Information and Systems

  IEICE Transactions on Information and Systems E94-D (10), 2015-2023, 2011

  The Institute of Electronics, Information and Communication Engineers

Keywords

• active learning
• query by committee
• LVCSR
• progressive alignment

Details

• CRID: 1390001204378833792
• NII Article ID: 10030193524
• NII Book ID: AA10826272
• DOI: 10.1587/transinf.e94.d.2015
• ISSN: 17451361
• ISSN: 09168532
• Web Site: http://t2r2.star.titech.ac.jp/cgi-bin/publicationinfo.cgi?q_publication_content_number=CTT100630434
• Web Site: http://www.jstage.jst.go.jp/article/transinf/E94.D/10/E94.D_10_2015/_pdf
• Text Lang: en
• Data Source:
  • JaLC
  • IRDB
  • Crossref
  • CiNii Articles
  • KAKEN
• Abstract License Flag: Disallowed