Атты І халықаралық конференция ЕҢбектері
жүктеу/скачать 8,57 Mb. Pdf көрінісі
|
Болатбек М. (1)
| Experiments
Acoustic Modeling An acoustic model was trained using CMU Sphinxtrain-1.0.8 [9]. The front-end module was set to output default parameters such as 13 mel-frequency cepstral coefficients with their first and second derivatives. Additionally, speaker adaptation techniques such as cepstral mean normalization [10], LDA [11] and MLLT [12] are performed on feature vectors. We used a context- dependent tied-state continuous Hidden Markov Model with 8 Gaussian mixtures per state [13]. The dictionary is compiled from the transcriptions and contains about 30000 words with their spellings as a phonetic transcription. It should be noted that there is still no consensus regarding the 234 Kazakh phonetic alphabet among the linguists [14]. Therefore, since the orthographic transcription of Kazakh roughly corresponds to a broad phonetic transcription, for the phoneme set a reduced form of the Kazakh alphabet is used, i.e. it includes those letters used in writing of Kazakh words. Also, for some letters there are variations in pronunciation depending on letter’s position or context in a word. Thus, letters, Е, О and Ө are pronounced as diphthongs in the beginning of a word. Letters Ю, Я are generally diphthongs except when used in the contexts CV and CVC, in such cases they obey vowel harmony and pronounced as their soft counterparts. Additionally, there is a SIL phone for silence. жүктеу/скачать 8,57 Mb. Достарыңызбен бөлісу:
|