Optimal use of trees in structural MAP adaptation for speaker verification

摘要

In speaker verification, the Structural Maximum-A-Posteriori(SMAP) adap­tation technique is often used to train speaker-adapted acoustic models by using available speech data in an efficient and flexible manner. In SMAP adaptation, a tree structure is used to represent the acoustic space of the human voice. We observed that one particular tree structure is not necessarily optimal for modeling the acoustic space of all speakers. In this paper, we propose a voting approach as a way to combine the decisions of multiple SMAP-adapted systems using different tree structures. We expect that this approach is more robust than using a single tree structure. We evaluate our proposed method on the 10sec4w-10sec4w task of NIST SRE 2006 and show that our method is more effective than the conventional SMAP adaptation as well as relevance MAP adaptation.