FIELD: computer equipment.
SUBSTANCE: method of creating a text-to-speech-text-to-speech (TTS) training set for teaching machine learning algorithm (MLA) to create machine-readable statements, which are input text, includes: obtaining a dictionary housing; parsing text representation of words from dictionary housing on phonemes; for each two selected words: creation of an artificial word based on combining separate phonemes from each of two words; making utterances from artificial words based on created two artificial words; obtaining estimates for created statements from artificial words, estimates are a parameter of human evaluation of the phoneme transition pattern near the tie point; determining acoustic properties of created utterances from artificial words; and classifier training based on created statements of artificial words, acoustic properties and estimates to create: quality parameter associated with new artificial word based on phonemes, wherein the quality parameter indicates whether a new artificial word based on phonemes naturally sounds.
EFFECT: creation of machine statements without user-visible delay.
16 cl, 6 dwg
