Font Size: a A A

Robustness Study Using Missing Feature Theory For Text-Independent Speaker Recognition

Posted on:2009-04-29Degree:DoctorType:Dissertation
Country:ChinaCandidate:W LuFull Text:PDF
GTID:1118360242495852Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
With the in-depth research on Speaker Recognition, it is important to improve the robustness of text-independent speaker recognition system in real environments. Due to the effective description of the distribution of speech database, statistical models, such as GMM (Gaussian Mixture Model), become the main technique in the area of text-independent speaker recognition. The performance of statistical models based speaker recogniton system relies on the similarity of training and testing environments, but the mismatch caused by background noise and channel distortion leads to dramatically performance decline in real environments.According to the fact that noise distorts speech differently in different time-frequency region and the redundancy in the speech signal, the thesis studies the detection and reconstruction of the highly corrupted features which called missing features in depth to reduce the mismatch of training and testing environments and improve the robustness of current speaker recogniton system, the thesis analyses background noise. The main content and results of study involved in this dissertation are divided into four parts:Firstly, the thesis studies the method of SS (spectral subtraction) based speaker recognition. The speech signal is enhanced by generalized SS, and then the MFCC features are extracted by the enhanced speech. The experiments on white noise and F16 cockpit noise in different SNRs (signal noise ratio) show that SS could improve the system performance in noisy environment to a certain extent. Further, the thesis demonstrates that it is difficult to get accurate estimates of the highly corrupted frequency bands only by SS-based speech enhancement method, which is the main reason of limiting the further development of speaker recognition in noisy environments.Secondly, according to the fact that noise distorts speech differently in different time-frequency region and the redundancy in the speech signal, the thesis proposes missing feature marginalization based on local SNR threshold method. Local SNR is used to identify different Mel-subbands features as 'reliable' or 'missing', only the reliable features are used to adapt the speaker models to calculate the output score. The method dramatically improve the system performance because that the highly corrupted features are removed from the inclusion in scoring. Based on the method, the thesis further presents a method of combined spectral subtraction and missing feature marginalization for robust speaker recogniton, and the system performance has a further improvement.Thirdly, based on the relevance of Mel-subbands in the same frame, the thesis presents two missing feature reconstruction methods based on statistical models: Cluster-based missing feature reconstruction and GMM-based missing feature reconstruction ,the former method has two steps, firstly clusters the training data, then uses Gaussian model to describe the distribution of each cluster separately; the latter mehod combines the two steps of the cluster-based method into one step in order to get more accurate description of the distribution of whole training feature sets. The experiments result shows that missing feature reconstruction method could get substantial increase in speaker recogniton system.Fourthly, based on the analysis that MFCC features describe the vocal track characteristics, but Pitch will effect the accurate description, the thesis proposes SMFCC (Smoothing MFCC) based speaker recognition. The experiments result shows that SMFCC feature is superior to MFCC feature, and the result in the female database, the SMFCC-based method has obvious advantages, furthermore the method has better robustness in time.
Keywords/Search Tags:Text-Independent
PDF Full Text Request
Related items