Author :
Songwatana, Kraisin ; Dejhan, Kobchai ; Miyanaga, Yoshikazu ; Khanthavivone, K.
Abstract :
Summary form only given. The paper presents a vowel recognition method for the Laotian spoken language. The vowels to be recognized consist of 12 pairs (short, long) of mixed and unmixed vowels ("a:,aa", "i:,ii", "/spl omega/:,/spl omega//spl omega/", "u:,uu", "e:,ee", "E:,EE", "o:,oo", "O:,OO", "/spl epsiv/:,/spl epsiv//spl epsiv/", "/spl omega/a:,/spl omega/a", "ia:,ia", "ua:,ua") and 3 additional vowels ("ai", "a/spl omega/", "am"). The vowel utterances are represented by an 18-dimensional vector of critical band intensities of the vocal tract transfer function. The vector of each frame is used as the basis for training a hidden Markov model and recognition of the vowels. A model for each vowel is generated and tests are then conducted from 7,020 words to evaluate the capability of such models. The results show an average recognition accuracy above 92%.
Keywords :
hidden Markov models; learning (artificial intelligence); speech recognition; transfer functions; vectors; Laotian language; bark scale; critical band intensity vector; hidden Markov model; vocal tract transfer function; vowel utterances; vowels recognition model; Hidden Markov models; Natural languages; Speech recognition; Testing; Transfer functions;
