UNCONTROLLED VOICE RESTORATION USING UNCONDITIONED DIFFUSION MODEL WITHOUT TEACHER Russian patent published in 2024 - IPC G10L21/00 

FIELD: computer technologies.

SUBSTANCE: invention relates to methods of processing and analyzing audio recordings, and can be used to improve the quality and intelligibility of speech recordings. Method of reconstructing voice in speech recordings comprises steps of: receiving audio data of a speech recording containing a voice audio signal; applying a probabilistic diffusion model, trained on the task of selecting a contribution function for noise reduction, for unconstrained generation of speech, wherein the probabilistic diffusion model is applied to the audio data of the speech recording in the form of a waveform containing random Gaussian noise; performing iterative sampling of the waveform using the conditional contribution function, which is the sum of the unconstrained contribution function, estimated by the probabilistic diffusion model, and logarithmic likelihood, to obtain a sample with a reduced amount of noise for the next iteration, until a denoised speech waveform is obtained; and outputting the processed voice audio signal containing the denoised speech signal waveform. Also disclosed is a system and a machine-readable medium for implementing the method.

EFFECT: high quality and legibility of speech recordings.

18 cl, 2 dwg, 5 tbl