A real-time pediatric dysarthria speech disorder detection using residual recurrent neural network with attention U-net based transformer encoder model

Speech disorders have a significant impact on quality of life, as they decrease the ability to define one's character, exercise autonomy, and frequently affect relationships and self-esteem, particularly in young children. Dysarthria is a neurological illness that affects motor speech pronunciation. Young children who experience this disorder have no issue with their understanding, but they have a problem expressing their words. They might struggle to communicate precisely and smoothly with their friends and family members due to this illness. A dysarthric child has significant trouble with communication, as this disorder causes poorly pronounced phonemes and poor speech articulation. To address this condition, numerous speech assistive technologies have been developed for consumers with dysarthria, tailored to the level of severity. Currently, deep learning (DL) systems offer potential for objective evaluation, thereby improving diagnostic accuracy. Its goal is to systematically analyze present approaches for detecting dysarthria based on severity levels. In this manuscript, a novel pediatric dysarthria disorder detection framework using residual recurrent neural network and transformer (PD3F-RRNNT) technique is proposed. The PD3F-RRNNT technique aims to develop a real-time recognition method for accurately detecting dysarthria speech disorders in children, supporting early diagnosis and intervention. Initially, the audio processing phase involved various steps, including voice activity detection (VAD), noise removal, pre-emphasis, framing, windowing, and normalization, to transform and extract significant data from audio signals. Furthermore, the PD3F-RRNNT method utilizes the transformer-attention-based U-Net (TransAttUnet) technique for feature extraction. Finally, the residual bidirectional gated recurrent unit (RBG) method is employed to detect and classify speech disorders accurately. The experimental validation of the PD3F-RRNNT model is performed under the dysarthria and non-dysarthria speech dataset. The comparison analysis of the PD3F-RRNNT model revealed a superior accuracy value of 99.50% compared to existing techniques.