LIP features for speech and speaker recognition

This paper implicitly differentiates between the quality of visual representation necessary for speech and speaker recognition and assesses the performance of visual lip features with respect to well established audio features. Blue lip highlighted data is used to show how variations in lip measurements can influence speech and speaker recognition. From these experiments and other researchers results [1] it is postulated that the fine detail of the lips is critical for speaker recognition, but conversely, the same amount of detail does not noticeably improve visual speech recognition. Visual error rates of 26.3% and 70% are achieved for cross-digit speaker and cross-speaker speech recognition respectively.