Improving Robustness of Speaker Recognition in Noisy and Reverberant Conditions via Training

Speaker recognition can be used as a security means to authenticate the speaker or as a forensic tool to determine who is likely to be the talker. For such critical applications, robustness or reliability of the system is crucial. In spite of the development and advancement in the field of speaker recognition, there are still many limitations and challenges. Amongst these, environment factors, in particular background noise and reverberation, are known to be difficult to tackle. Environmental noises and reverberation compromise the accuracy of speaker recognition, especially when the signal to noise ratio (SNR) becomes low and reverberation time is long. Noises and reverberation mitigate reliability of speaker recognition systems via signal transmission channel mismatch. This paper is presented from attempts to improve system robustness by adding noises and convoluting room impulse responses in the training phase of typical Gaussian Mixture Model based speaker recognition systems. Validation tests were carried with emulated noisy and reverberant conditions with controlled signal to noise ratios and reverberation times. Two scenarios have been considered the first one used the clean speech samples in enrolment phase and the second included noisy or reverberant samples in enrolment phase, thus the potentials and limitations of including noisy and reverberant samples in the training phase to improve system robustness is identified.