Font Size: a A A

Noise Immunity Of Continuous Speech Recognition Research

Posted on:2007-04-06Degree:MasterType:Thesis
Country:ChinaCandidate:J LinFull Text:PDF
GTID:2208360185956192Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Because noise causes the mismatch between the acoustic models and the testing speech, the performance of speech recognition systems will degrade rapidly in noisy environments. Therefore, noise robust technology is a very crucial problem for the speech recognition.In this thesis, first we analyzed and designed a traditional continued speech recognition system, which based on HMM and MFCC speech features. Then we researched some noise robust technologies based on that system.A number of techniques have been developed to reduce the mismatch caused by environment noise over the past decades. They include: 1) speech enhancement, 2) extracting robust speech features, 3) speech model compensation for noisy environments, 4) missing feature. In this thesis, our research is forced on speech enhancement method and probabilistic union model based on missing-feature. Through analyzing the disadvantages of tradition Wiener filter and adaptive noise canceller, we proposed two new methods for enhancing speech, which are Improved Iteration Wiener filter method and Improved CTRANC algorithm. The results show that the new methods are an effective scheme improving the ability of removing noise. In addition, we proposed the Posterior Union Model (PUM), which improves over the conditional union model by retaining the advantage of requiring no identity of the noisy components, and by additionally offering a means of optimally estimating the model order, therefore enhancing the capability of the model for dealing with nonstationary noise.However, this PUM is only effective for partial noise corruption. It lacks robustness for wide-band noises corruption. To cope with the problem, we proposed to...
Keywords/Search Tags:speech recognition, robustness, noise, HMM, speech enhancement, PUM
PDF Full Text Request
Related items