Robust Speaker Recognition Using Both Vocal Source and Vocal Tract Features Estimated from Noisy Input Utterances

Motivated by the mechanism of speech production, we present a novel idea of using source-tract features in training speaker models for recognition. By considering the severe degradation occurring when a speaker recognition system operates under noisy environment, which could well be due to the missing of speaker-distinctive information, we propose a robust feature estimation method that can capture the source and tract related speech properties from noisy input speech utterances. As a simple yet useful speech enhancement technique, spectral subtractive-type algorithm is employed to remove the additive noise prior to feature extraction process. It is shown through analytical derivation as well as simulation that the proposed feature estimation method leads to robust recognition performance, especially for very low signal-to-noise ratios. In the context of Gaussian mixture model-based speaker recognition with the presence of additive white Gaussian noise in the input utterances, the new approach produces consistent reduction of both identification error rate and equal error rate at signal-to-noise ratios ranging from 0 dB to 15 dB.