EXTRACTION AND MODELING CONTEXT DEPENDENT PHONE UNITS FOR IMPROVEMENT OF CONTINUOUS SPEECH RECOGNITION ACCURACY BY PHONEMES CLUSTERING