Audio-Visual Speech Recognition Using New Lip Features Extracted from Side-Face Images

This paper proposes new visual features for audio-visual speech recognition using lip information extracted from side-face images. In order to increase the noise-robustness of speech recognition, we have proposed an audio-visual speech recognition method using speaker lip information extracted from side-face images taken by a small camera installed in a mobile device. Our previous method used only movement information of lips, measured by optical-flow analysis, as a visual feature. However, since shape information of lips is also obviously important, this paper attempts to combine lip-shape information with lip-movement information to improve the audio-visual speech recognition performance. A combination of an angle value between upper and lower lips (lip-angle) and its derivative is extracted as lip-shape features. Effectiveness of the lip-angle features has been evaluated under various SNR conditions. The proposed features improved recognition accuracies in all SNR conditions in comparison with audio-only recognition results. The best improvement of 8.0% in absolute value was obtained at 5dB SNR condition. Combining the lip-angle features with our previous features extracted by the optical-flow analysis yielded further improvement. These visual features were confirmed to be effective even when the audio HMM used in our method was adapted to noise by the MLLR method.