To implement the project, a machine learning model that analyzes recorded speech in the form of audio files and identifies pronunciation defects in children, was used. This allows to hold initial consultation and make diagnosis. The Spectrogram widget looks like a test in a chatbot that can record audio files, and the child is tested in the form of games.
Currently the neural network automatically determines the probability that there is a speech defect in the audio recording and presents the corresponding information to the speech therapist in the administrator panel to speed up the process of checking the audio recordings.
Several hundred testing questionnaires have already been collected with the help of the widget, and the accuracy of speech defects recognition is more than 80% for high-quality audio. In the future, the plans include a complete transition from the human-in-the-loop model, where a person’s participation in diagnostics is necessary, to the complete automation of the process with the help of a retrained model.