METHOD AND SERVER FOR CONVERTING TEXT TO SPEECH Russian patent published in 2022 - IPC G10L13/08 

FIELD: computing technology.

SUBSTANCE: technical result is achieved by generating a signal segment at the current iteration, involving storing first tensor data calculated by the first hidden layer at this iteration and containing tensor fragment data. The tensor fragment data is used at the current iteration to generate a signal segment and should be used at the next iteration to generate the next signal segment. Due to the generation of the next signal segment at the next iteration, involving storing second tensor data calculated by the first hidden layer at the next iteration, the second tensor data does not contain redundant tensor fragment data identical to the tensor fragment data from the first tensor data.

EFFECT: increased response speed of the IPA system (reduced time between the speech fragment of the user and the provision of at least part of the speech fragment generated by the computer) due to the optimisation of the amount of operations and the size of memory needed for the CNN network to form signal segments in real time.

24 cl, 8 dwg