Sensory integration in audiovisual automatic speech recognition
Peter L. Silsbee · 2002
Methods of integrating audio and visual information in an audiovisual HMM-based ASR system are investigated. Experiments involve discrimination of a set of 22 consonants, with various integration strategies. The role of the visual subsystem is varied; for example, in one run, the subsystem attempts to classify all 22 consonants, while in other runs it attempts only broader classifications. In a second experiment, a new HMM formulation is employed, which incorporates the integration into the HMM at a pre-categorical stage. A single variable parameter allows the relative contribution of audio and visual information to be controlled. This form of integration can be very easily incorporated into existing audio-based continuous speech recognizers.>