Machine recognition of Hindi consonants and distinctive features using vector quantization

Abstract

This paper describes the results of experiments conducted to test the efficiency of the VQ algorithms to recognize and distinguish minute differences among consonant features of Hindi. The database was created by combining all the frequently occurring 29 Hindi consonants with vowel /a/ to form CV-type syllables and recorded by 20 male Hindi speakers. The acoustic features in the form of cepstral vectors were derived for each syllable file in the training set. For each syllable, the set of 20 codebooks for 20 speakers were first clubbed to form the universe for that syllable. The LBG algorithm was used to generate the codebook of size 64. Using the codebooks as reference templates the recognition rate was found to be 73.28% out of the total of 580 (29*20) test samples. Further analysis was carried out to see how different consonants were recognized when they were clubbed into different groups according to the manner and place of their articulation. The average recognition rate for the consonants corresponding to the same place of articulation and the same manner of articulation was found to be 83.38% and 76.25%, respectively. Results of detailed distinctive feature analysis carried out to find the differences in the rate of recognition are also discussed.