Audio-Visual Speech Analysis & Recognition

Human speech production and perception mechanisms are essentially bimodal. Interesting evidence for this audiovisual nature of speech is provided by the so-called Mc Gurk effect. To properly account for the complementary visual aspect we propose a unified framework to analyse speech and present our related findings in applications such as audiovisual speech inversion and recognition. Speaker's face is analysed by means of Active Appearance Modelling and the extracted visual features are integrated with simultaneously extracted acoustic features to recover the underlying articulator properties, e.g., the movement of the speaker's tongue tip, or recognize the recorded utterance, e.g. the sequence of the numbers uttered. Possible asynchrony between the audio and visual stream is also taken into account. For the case of recognition we also exploit feature uncertainty as given by the corresponding front-ends, to achieve adaptive fusion. Experimental results are presented in QSMT, MOCHA and CUAVE audiovisual databases.