Visual Speech Recognition

In recent years, Visual speech recognition has a more concentration, by researchers, than the past. Because of the leakage of the visual processing of the Arabic vocabularies recognition, we start to search in this field. Audio speech recognition concerned with the acoustic characteristic of the signal, but there are many situations that the audio signal is weak of not exist, and this will be a point in Chapter 2. The visual recognition process focuses on the features extracted from video of the speaker. These features are to be classified using several techniques. The most important feature to be extracted is motion. By segmenting motion of the lips of the speaker, an algorithm has manipulate it in such away to recognize the word which is said. But motion segmentation is not the only problem facing the speech recognition process, segmenting the lips itself is an early step in the speech recognition process, so, to segment lips motion we have to segment lips first, a new approach for lip segmentation is proposed in this thesis. Sometimes, motion feature needs another feature to support in recognition the spoken word. So in our thesis another new algorithm is proposed to use motion segmentation by using the Abstract Difference Image from an image series, supported by correlation for registering images in the image series, to recognize ten words in the Arabic language, the words are from “one” to “ten” in Arabic language. The algorithm also uses the HU-Invariant set of features to describe the Abstract Difference Image, and uses a three different recognition methods to recognize the words. The CLAHE method as a filtering technique is used by our algorithm to manipulate lighting problems. Our algorithm based on extracting the differences details from a series of images to recognize the word, achieved an overall results 55.8%, it is an adequate result for our algorithm when integrated in an audio-visual system