Text-independent speaker indentification in a distant-talking multi-microphone environment = 원거리 다채널 환경에서 문맥독립 화자식별

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 308
  • Download : 0
In this dissertation, we propose new combination schemes to integrate the identification results obtained by recognizing the speech inputs recorded simultaneously with the individual microphones with the aim of achieving the best possible speaker identification rate. Recent speaker identification technologies perform reasonably well when speech signals are captured in noise-free environments using close-talking microphones. However, such ideal acoustic conditions are generally unrealistic and the speaker identification rate could be significantly degraded due to a variety of causes. In order to deal with such problems, we propose a likelihood-based combination method to integrate the identification results by rescoring the average log-likelihood of the hypothesis. And we upgrade the existing combination schemes by applying weight to the individual identification result based on the frame``s relative entropy. In many practical situations, the likelihood scores lie within different dynamic ranges since there exists a score variability, which could come from different information, sound sources, etc. Thus, the likelihood scores themselves do not necessarily reflect their respective significances exactly. The speaker identification task generally does not require rescoring since the decision is made using the likelihood score from a single utterance. In this dissertation, however, the average log-likelihoods are rescored in order to make them reflect their respective significances, and they are used to combine the identification results for performance improvement in speaker identification. In addition, we propose a combination method to fuse the identification results softly by measuring the degree of the confidence in their respective classification results. From the experimental results, it is demonstrated that the proposed combinations greatly enhance the identification performances in a noisy distant-talking environment. The experimental result shows tha...
Advisors
Kim, Hoi-Rinresearcher김회린researcher
Description
한국정보통신대학교 : 공학부,
Publisher
한국정보통신대학교
Issue Date
2008
Identifier
393011/225023 / 020025354
Language
eng
Description

학위논문(박사) - 한국정보통신대학교 : 공학부, 2008.6, [ x, 95 p. ]

Keywords

Speaker Identification; Multiple microphone; Speaker Recognition; 화자인식; 화자식별; 다채널 환경

URI
http://hdl.handle.net/10203/54607
Link
http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=393011&flag=dissertation
Appears in Collection
School of Engineering-Theses_Ph.D(공학부 박사논문)
Files in This Item
There are no files associated with this item.

qr_code

  • mendeley

    citeulike


rss_1.0 rss_2.0 atom_1.0