LANGUAGE IDENTIFICATION USING ACTIVE APPEARANCE MODELS

Abstract

I anguage identification is an area of speech processing where the language being spoken is identified automatically. The problem of automatic language identification has been studied extensively with audio speech and its performance has been evaluated satisfactorily. 1-lowever visual language identification, which focuses on the identification of the spoken language using only the features available in video, has not met with as much success. Visual language identification can be an effective tool in law enforcement and even as a component of an audio visual speech recognition system. This can also be crucial to identification of the language in conditions when the audio may not be available or is of low quality due to a large amount of noisc in the video. This serves as motivation for this work. In this thesis, the methods that have already been applied effectively in the fiekl of audio language identification are reviewed and a new method is proposed. Before looking at methods that have already been applied to visual language identification, active appearance models that have been used extensively in these works are examined. A model of viseme representation is chosen to allow a simple classifier based separation of the language. Standard models require the information available from a languages model in order to be able to identify the language. However the language model information of all languages may not be available to us. I-hence this work focuses on language classification without using such a model. Also the principal components of the active appearance models are examined as features for such a classifier and their performance is evaluated