RECOGNIZING OF MIXED SPEECH Russian patent published in 2019 - IPC G10L15/06 G10L15/16 G10L15/20 

FIELD: data processing.

SUBSTANCE: invention relates to means of recognizing mixed speech. First neural network is trained to recognize a speech signal pronounced by a speaker with a higher level of speech characteristics from a sample of mixed speech. Second neural network is trained to recognize a speech signal pronounced by a speaker with a lower level of speech characteristics from a sample of mixed speech. Mixed speech sample is decoded by a first neural network and a second neural network by optimizing the combined probability of observing said two speech signals, where the combined probability means the probability that a particular frame is a switching point of the speech characteristic. Third neural network is taught to predict switching of the speech characteristic. Mixed speech sample is decoded based on said prediction.

EFFECT: high accuracy of recognizing mixed speech.

15 cl, 5 tbl, 6 dwg