Font Size: a A A

The Research On Channel Robustness Of Text-independent Speaker Verification

Posted on:2010-11-17Degree:MasterType:Thesis
Country:ChinaCandidate:J S ChaoFull Text:PDF
GTID:2178360275470077Subject:Biomedical engineering
Abstract/Summary:PDF Full Text Request
Speaker verification is a sort of biometrics technology which aims to determine whether a person's claimed identity is correct or whether the person is an imposter. This technology is widely used in many identity validation fields such as entrance guard system, telephone banking, database accessing. The speaker verification has achieved satisfactory results in clean speech environment. However, those systems'performance decreases drastically under the channel or noise environment.Therefore this thesis studies on text-independent speaker verification and channel robustness methods on feature level, model level and system level. It mainly focuses on the following aspects: 1) Analyze several speech feature characteristics and extraction algorithms. Implement useful robust algorithms on the frame level. 2) Based on the current successful UBM-GMM system, the excellent channel robust algorithm - Latent Factor Analysis (LFA) is applied to the system, which gains great performance. 3) The thesis explains basic theory of SVM, and then introduces it to speaker verification. The latest channel compensation method-Nuisance Attribute Projection (NAP) is also examined in our best supervector SVM system. 4) In order to improve the performance of text-independent speaker verification, the thesis works out some fusion methods of sub-systems. It proposes a method which combines MFCC based system and Pitch based system to gain better result. Last we combine all the sub-systems to yield the best robust system of speaker verification.The project makes use of speech database and performance measure criteria from National Institute of Standards and Technology (NIST) Speaker Recognition Evaluation (SRE). This standard speech database was recorded via a number of telephone, mobile and microphone channels. The experiments show validity of those channel robust algorithms so that our systems could be used in real circumstances.
Keywords/Search Tags:speech feature, speaker verification, Gaussian Mixture Model, Support Vector Machine, channel robustness
PDF Full Text Request
Related items