The method for generating a training set of voice commands based on speech synthesis technology

The advent of modules based on neural network technologies for recognizing audio commands makes it much easier to solve problems related to remote control of various devices designed for a wide range of users. At the same time, neural network modules can work offline and do not require a network connection to process audio information using other computing devices. However, the memory size of these modules is limited, and the probability of recognizing audio commands directly depends on the size of the training sample with a set of samples of executed commands. Accordingly, there is a scientific problem associated with increasing the probability of recognition by neural network modules of voice commands.

A method has been developed for generating a limited-size training sample of synthesized voice commands with a variety of pronunciation of words that differ in frequency, duration, phoneme composition, and stress in a word, which will increase the probability of recognition by neural network modules of voice commands for an unlimited range of users.

A comparative analysis of the probability of recognizing voice commands before and after training a neural network module using a training sample of synthesized voice commands is carried out. In the course of the conducted research, it was found that modern neural network modules are able to store in memory about 10 groups of 8 models of voice commands each. Accordingly, the probability of recognizing each group of voice commands will be about 80%.

Due to the use of a synthesized sample of voice commands, the number of potential users has increased, including those with speech defects, who can use such modules without additional training. The use of synthesized speech models for training neural network modules makes it possible to automate the learning process, reduce their training time, and significantly increase the likelihood of these modules recognizing voice commands. Further improvement of speech recognition technology will reduce the likelihood of false recognition of voice commands and make this technology available to a wide range of users.