A Two-Stage Algorithm for Multi-Speaker Identification System

In this paper, a two-stage multi-speaker identification (SID) system is proposed for mixed speeches with multiple speakers speaking simultaneously. By investigating the second stage processing, we improved the performance of multi-speaker SID from 94.6% to 99.0% on a standard testing set, and comparing with another state-of-art system, the proposed results were also a little better. We also examined the configure parameters of proposed algorithm, and found that the gain compensation parameter and composition model were crucial for multi-speaker SID. Also, the likelihood constrained parameter was an important improvement compared with conventional SID.