Speaker-Dependent Voice Activity Detection Robust to Background Speech Noise

In this paper, we proposed a speaker-dependent voice activity detection (VAD) algorithm that only extracts the speech period uttered by a target user. Based on our survey of the recognition error of real speech data collected in “VoiceTra,” which is a speech-to-speech translation system for smartphones, we found that many word insertion errors are caused by background speakers’ speech. Our VAD, which consists of three GMMs (noise and speaker-independent GMMs as used in a traditional GMM-based VAD and a speaker-adapted GMM) can be used for speech detection of the target speaker. In the VAD evaluations for the test utterances with background speakers’ speech, our proposed VADs achieved better performance than the conventional VAD. Also speech recognition experiments demonstrated that an ASR system with our proposed VAD achieved better performance than an ASR system using the conventional VAD.