Phone-based speaker verification with various adaptation configurations

This paper introduces an examination of several adaptation schemes for phone-based speaker verification. It is argued that different phonemes convey different amount of valuable information about speaker classification. Thus, the presented study tries to find more optimal settings to exploit this information. Experiments with short duration of training and testing of clean and telephone text-independent speech highlighted the superiority of one configuration over the rest. This configuration implements a two-stage adaptation of phone models, outperforming the standard phone-independent GMM-based system. The paper also considers adaptation of a subset of the whole phonetic set, and its comparative improvement of performance. Experiments showed that this partial adaptation approach, beyond reducing storage requirement, significantly improves the overall performance.