Employment of Spectral Voicing Information for Speech and Speaker Recognition in Noisy Conditions
Peter Janovi, Mnevver Kker · InTech eBooks · 2008
This chapter described our recent research on representation and modelling of speech signals for automatic speech and speaker recognition in noisy conditions. The chapter consisted of three parts. In the first part, we presented a novel method for estimation of the voicing information of speech spectra in the presence of noise. The presented method is based on calculating a similarity between the shape of signal short-term spectrum and the spectrum of the frame-analysis window. It does not require information about the F0 and is particularly applicable to speech pattern processing. Evaluation of the method was presented in terms of false-rejection and false-acceptance errors and good performance was demonstrated in noisy conditions. The second part of the chapter presented an employment of the voicing information into the missing-feature-based speech and speaker recognition systems to improve noise robustness. In particular, we were concerned with the mask estimation problem for voiced speech. It was demonstrated that the MFT-based recognition system employing the estimated spectral voicing information as a mask obtained results very similar to those of employing the oracle voicing information obtained based on full apriori knowledge of noise. The achieved results showed significant recognition accuracy improvements over the standard recognition system. The third part of the chapter presented an incorporation of the spectral voicing information to improve modelling of speech signals in application to speech recognition in noisy conditions. The voicing-information was incorporated within an HMM-based statistical framework in the back-end of the ASR system. In the proposed model, a voicing-probability was estimated for each mixture at each HMM state and it served as a penalty during the recognition for those mixtures/states whose voicing information did not correspond to the voicing information of the signal. The evaluation was performed in the standard model and in the missing-feature model that had compensated for the effect of noise and experimental results demonstrated significant recognition accuracy improvements in strong noisy conditions obtained by the models incorporating the voicing information.