An acoustic study of syllable onsets : a basis for Thai continuous speech recognition system

Abstract

To develop a new acoustic speech units on modelling of the Thai onset units. The concept of onset and rhyme units is applied to Thai continuous speech recognition. Thai syllables are acoustically analysed and found that a syllable is composed of a pair of onset and rhyme units. The onset unit incorporates transitional period existed between releasing consonant and its adjacent vowel. The rhyme unit covers both vowel and arresting consonant. The transitional period has unique acoustic characteristics depending on releasing consonant and vowel which is crucial in recognition of the consonant. Two acoustic models of the onset-rhyme unit are introduced in this dissertation-Phonotactic Onset-Rhyme Models (PORMs) and Contextual Onset-Rhyme Models (CORMs). The PORMs consider the same releasing consonant in different context as different models. This results in 792 onset units and 200 rhyme units. The CORMs consider the same releasing consonant within similar short-long vowel context as the same models. The number of onset units is then reduced to only 297 units. In modelling of the onset units, the onset unit overlaps over vowel seqment of rhyme unit to cover transitional period between releasing consonant and adjacent vowel. Two overlapping techniques are proposed in modelling of the onset units-fixed duration overlap and variable duration overlap. The fixed duration overlap has constant duration at 10, 20, or 30 ms into the vowel segment. The variable duration overlap has variable duration at either 5%, 10%, 15%, 20%, or 25% of the vowel segment. Due to limited speech data and only one male speaker, only partial set of the onset and rhyme units are modelled in the speaker-dependent recognition system. The PORMs and CORMs contain 384 onset units and 218 onset units, respectively. Both models share the same 144 rhyme units. Acoustic models of these onset and rhyme units are created using Hidden Markov Models. The lowest onset error rate achieved is 10.38% using the CORMs at 25% overlap. Considering both onset and rhyme units, the PORMs provide better word error rate at 13.53% using 20% overlap with no grammar. The CORMs give out 16.51% word error rate at 15% overlap. The phone models give out 37.12% word error rate. Hence, the PORMs outperform the phone models up to 55.76% reduction in word error rate