Font Size: a A A

Research On New Method Of Rubust Speech Recognition

Posted on:2006-04-28Degree:MasterType:Thesis
Country:ChinaCandidate:S F LinFull Text:PDF
GTID:2168360152975512Subject:Detection Technology and Automation
Abstract/Summary:PDF Full Text Request
Noisy speech recognition is one of the active research areas of speech signal processing. For the difference between the training condition and real world, especially all kinds of noise buried in the speech signals, the performances of most recognition systems are degraded greatly. Thus so far, few speech recognition systems are really put to use.This paper starts from the speech models, and main works focus on developing new noisy speech recognition methods. The influences of noise on the performances of speech recognition are thoroughly analyzed and compared. Then two new methods are presented. The LVQ/DTW method, Learning Vector Quantization (LVQ) technique combined with the Dynamic Time Warping (DTW) algorithm, is the first hybrid model. In this model, the Dynamic Time Warping algorithm is used as the front-end of Learning Vector Quantization network, warping the epoch of the input utterances. The task of classification and recognition is completed by the Learning Vector Quantization network, which is modified in the learning algorithm. Compared with traditional methods, the training and testing process for the LVQ/DTW model is very simple. The experimental results demonstrate that recognition rate can be improved from 26% to 50% when signal and noise ratio (SNR) is 10dB.The second method fuses the modeling ability in time domain of hidden markov model and the powerful classification and decision ability of wavelet neural network. Compared with the LVQ/DTW method, this method has good performance in noisy speech condition. The recognition rate is 66% when the SNR is lOdB. Experiments also show that this method is fit with the low SNR condition.Meanwhile, speech enhancement techniques are becoming another active research area in speech processing. Speech enhancement techniques are applied to speech recognition. Noisy speech signal are processed to improve its SNR before recognition. Two methods are considered, spectrum subtraction technique and wavelet denoising techniques. SNR of the original noisy speech can be improved at least 50%. The recognition rate can reach 80% when SNR is lOdB. Furtherexperiments show that the performance of this method based on speech enhancement is good in noisy speech.
Keywords/Search Tags:speech recognition, noise, learning vector quantization, wavelet network, speech enhancement
PDF Full Text Request
Related items