Phone adaptive training for short-duration speaker verification

Phone adaptive training (PAT) aims to derive a new acoustic feature space in which the influence of phone variation is minimised while that of speaker variation is maximised. Originally proposed in the context of speaker diarization, our most recent work showed the utility of PAT in short-duration, automatic speaker verification where phone variation typically degrades performance. New to this contribution is the assessment of PAT utilising automatically generated acoustic class transcriptions whose number is controlled by regression tree analysis. Experimental results using a standard database show that PAT delivers significant improvements in the performance of a state-of-the-art iVector speaker verification system.