Robust Extraction of Desired Speaker's Utterance in Overlapped Speech

摘要

In this paper, we propose a speaker indexing method using speaker verification technique to extract one desired speaker's utterances from conversational speech. To solve the overlapped speech problem, we construct overlapped speech models with the observed conversational speech itself. The overlapped speech models include overlapped speech of target and cohort speaker, and speech model of two cohort speakers. In order to evaluate the proposed method, we made a simulated conversational speech that has up to 50% overlapping segments. The equal error rate was reduced by up to 43.7% compared with the conventional methods that use a target speaker model only, and use a target model and overlapped speech model trained with a speaker independent of large speech database.