Assessment of syllable intelligibility based on convolutional neural networks for speech rehabilitation after speech organs surgical interventions

Abstract

Head and neck cancer patients often have side effects that make speaking and communicating more difficult. During the speech therapy the approach of perceptual evaluation of voice quality is widely used. First of all, this approach is subjective as it depends on the listener’s perception. Secondly, the approach requires the patient to visit a hospital regularly. The present study is aimed to develop the automatic assessment of pathological speech based on convolutional neural networks to give more objective feedback of the speech quality. The structure of the neural network has been selected based on experimental results. The neural network is trained and validated on the dataset of phonemes which are represented as Mel-frequency cepstral coefficients. The neural network is tested on the syllable dataset. Recognition of the phoneme content of the syllable pronounced by a patient allows to evaluate the progress of the rehabilitation. A conclusion about the applicability of this approach and recommendations for the further improvement of its performance were made.