Voice activity detection (VAD) is one of the most critical issues on performance degradation of speech recognition in noisy environment applications. A real-time VAD was developed by using face parameters (eye and lip contours) as a front-end for the traditional speech and noise (audio) GMMbased method. Speech recognition performance of the audiovisual VAD is shown to be comparable with audio-only VAD, for a shopping mall background noise. Advantages and limitations of introducing the visual information are discussed.

Index Terms: voice activity detection, audio-visual, speech recognition, noisy environment, real-time.

@inproceedings{ishi10_avsp,
  title     = {Real-time audio-visual voice activity detection for speech recognition in noisy environments},
  author    = {Carlos T. Ishi and Miki Sato and Norihiro Hagita and Shihong Lao},
  year      = {2010},
  booktitle = {Auditory-Visual Speech Processing},
  pages     = {paper P5},
}