Fusion of cohort-word and speech background model based confidence scores for improved keyword confidence scoring and verification

This paper presents results for Keyword Verification (KV) experiments using Cohort-Word Verification (CWV) and neural-network fusion of CWV and Speech Background Model based Verification (SBMV). Baseline experiments found that CWV excelled in short-word KV while SBMV was more robust for long-word KV. Fusion of CWV and SBMV yielded dramatic improvements in both short-to-medium length KV using the fusion of two CWV systems, and medium-to-long length KV using fusion of SBMV and CWV. Overall, through target word phone length dependent fusion of verifiers, it was possible to at least halve the false rejection rate of the baseline SBMV verifier for all evaluated keyword length classes.