Toward Clustering Persian Vowel Viseme: A New Clustering Approach based on HMM

This paper sorts out the problem of Persian Vowel viseme clustering. Clustering audio-visual data has been discussed for a decade or so. However, it is an open problem due to shortcoming of appropriate data and its dependency to target language. Here, we propose a speaker-independent and robust method for Persian viseme class identification as our main contribution. The overall process of the proposed method consists of three main steps including (I) Mouth region segmentation, (II) Feature extraction, and (IV) Hierarchical clustering. After segmenting the mouth region in all frames, the feature vectors are extracted based on a new look at Hidden Markov Model. This is another contribution to this work, which utilizes HMM as a probabilistic model-based feature detector. Finally, a hierarchical clustering approach is utilized to cluster Persian Vowel viseme. The main advantage of this work over others is producing a single clustering output for all subjects, which can simplify the research process in other applications. In order to prove the efficiency of the proposed method a set of experiments is conducted on AVAII.