Test token driven acoustic balancing for sparse enrollment data in cohort GMM speaker recognition

摘要

In this study, we address the problem of sparse train/test data for in-set/out-of-set speaker recognition. Sparse enrollment data presents a unique challenge due to a lack of acoustic space coverage. The proposed algorithm focuses on filling acoustic holes and fortifying the acoustic information using the claimed speaker's test token histogram. This scheme is possible by using a GMM model to classify the speaker phone information at the feature level. Parallel GMM training with EM using the most occurring (top) and least occurring (bottom) acoustic feature is called “Top-Down Bottom-Up (TDBU)”, and the method employing the acoustic token histogram of test token using the TDBU is called “TDBU using Test Token Histogram (TTH)”. Since TTH provides test data histogram information, the most occurred (top) parts in test data fortify the its discriminating ability using same acoustic tokens in enrollment data. The less occurred (bottom) part in test data provide acoustic hole information so that the mismatched acoustic hole between enrollment and test data can be filled in chance. The TDBU-TTH method is evaluated using telephone conversation speech from the FISHER corpus with 5 second train sets. The TDBU-TTH improves on average 2.17% absolute EER over the TDBU, and an average 4.03% absolute EER improvement over GMM-UBM baseline using 2 second test data. The proposed algorithm improvement is a noteworthy stage to compensate for both sparse enrollment data and limited test data.