A segment-based audio-visual speech recognizer: data collection, development, and initial experiments