The paper addresses problem of improvement of nasal and sonorant discrimination in different vowels environments. Both selection of more appropriate features for discrimination of phonemes and more advanced training has been investigated. Selection of more appropriate features has been performed by checking different signal analysis algorithms (comparing filterbank analysis and cepstrum procedures) and combining different acoustic events describing features (evaluation of coarticulation). Advanced pattern classification algorithms such as scaled rotation maximization and regularized discriminant analysis were applied to phoneme discrimination task also. It has been shown that recognition accuracy depends strongly on the acoustic environment of consonant. Best results were observed in the environment of open vowel a while closed vowel i is more complicated one. The evaluation of these factors allowed to reduce discrimination error rate from 25-30 % to 6.6% in the context ofopen vowel a and 12.1% in the context of closed vowel i.
This work is licensed under a Creative Commons Attribution 4.0 International License.