Robust speaker recognition in presence of non-trivial environmental noise (toward greater biometric security)

• Main Author: Al-Noori, A. H. Y.
• Published: University of Salford 2017
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.736442

The aim of this thesis is to investigate speaker recognition in the presence of environmental noise, and to develop a robust speaker recognition method. Recently, Speaker Recognition has been the object of considerable research due to its wide use in various areas. Despite major developments in this field, there are still many limitations and challenges. Environmental noises and their variations are high up in the list of challenges since it impossible to provide a noise free environment. A novel approach is proposed to address the issue of performance degradation in environmental noise. This approach is based on the estimation of signal-to-noise ratio (SNR) and detection of ambient noise from the recognition signal to re-train the reference model for the claimed speaker and to generate a new adapted noisy model to decrease the noise mismatch with recognition utterances. This approach is termed “Training on the fly” for robustness of speaker recognition under noisy environments. To detect the noise in the recognition signal two different techniques are proposed: the first technique including generating an emulated noise depending on estimated power spectrum of the original noise using 1/3 octave band filter bank and white noise signal. This emulated noise become close enough to original one that includes in the input signal (recognition signal). The second technique deals with extracting the noise from the input signal using one of speech enhancement algorithm with spectral subtraction to find the noise in the signal. Training on the fly approach (using both techniques) has been examined using two feature approaches and two different kinds of artificial clean and noisy speech databases collected in different environments. Furthermore, the speech samples were text independent. The training on the fly approach is a significant improvement in performance when compared with the performance of conventional speaker recognition (based on clean reference models). Moreover, the training on the fly based on noise extraction showed the best results for all types of noisy data.