METHODS OF TRAINING MODEL AND VOICE CONVERSION AND DEVICE, OTHER DEVICE AND DATA MEDIUM Russian patent published in 2024 - IPC G10L13/02 G10L15/18 

FIELD: computer engineering.

SUBSTANCE: invention relates to computer engineering for processing audio data. Result is achieved by steps of obtaining a first voice characteristic from any audio sample fragment using a voice extraction network in an initial voice conversion model; obtaining a first semantic characteristic based on a first voice characteristic and a linear spectrogram corresponding to the audio data sample using a voice removal network in the original voice conversion model, wherein the first semantic characteristic is a characteristic of the audio data sample, which is not associated with the speaker’s voice, but with semantic information; obtaining synthesized audio data based on a first semantic characteristic and a second voice characteristic of the target audio data corresponding to the audio data sample using a vocoder in the original voice conversion model; and obtaining a trained voice conversion model by training the initial voice conversion model based on the target audio data and the synthesized audio data corresponding to each audio data sample fragment.

EFFECT: high accuracy of content information of an initial audio signal during voice conversion.

18 cl, 8 dwg