Reduced sets of subword units for continuous speech recognition of Portuguese

摘要

An investigation is presented concerning two sets of subword units for continuous speech recognition, which are based on the characteristics of the Portuguese language. In the first set, with 149 units, it is considered that syllables which contain an epenthetic vowel can be formed by two CV (consonant-vowel) units. In the second set, with 254 units, these syllables are regarded as CCV (consonant-consonant-vowel) units. The recognition performance obtained with the two unit sets are comparable if a bigram model is used at the unit level. In this case, the first set is definitely preferable because it enables the complexity to be significantly reduced.