A robust sparse auditory feature for speaker verification

We propose a novel robust feature extraction technique for speaker verification based on the human auditory system and sparse representation. In the technique, considering there is a small number of active neurons in the primary auditory cortex, the acoustic stimulus can be encoded and the sparse independent events can be used to represent the characteristics of the neurons, where every dictionary is learned from every speaker training sample, so that it has more individual information of the speaker and is useful for discriminating different speakers with less dictionary atoms. Moreover, the atom responses are similar to the neurons reacting on stimulus and the occurrence frequency of each atom in dictionaries is adopted as the feature for speaker verification. The robustness of the new technique is better in terms of a strategy to represent natural sounds. The experimental results show that the new technique system outperforms the baseline system on two typical corpuses.