A novel speech signal feature extraction technique to detect speech impairment in children accurately

Speech signal processing and extracting useful information from speech signal is necessary for speech language impairment (SLI) detection in children. Although different features has been suggested for SLI detection, there is still a scope exist for exploration of other methods. A comparative study of different techniques for feature extraction can be done to find the optimal feature extraction technique. In this work, a study has been carried out to obtain optimal feature extraction technique for SLI detection. Inputs used for SLI detection here are the speech signals recorded from children. Features are first extracted from the recorded speech signals using various feature extraction techniques. The feature extraction techniques that has been implemented are relative spectral transform - perceptual linear prediction (RASTA), wavelet packet transform (WPT), linear predictive coding (LPC), perceptual linear prediction (PLP), Mel-Frequency cepstral coefficients (MFCC), complex quantization cepstral coefficient (CQCC), perceptual noise cepstral coefficients (PNCC). The features extracted are then given to deep learning models namely transformer, temporal convolutional networks (TCN) and TabNet for SLI detection. The result obtained has highest accuracy of 100.00 % using PNCC feature combined with TabNet method. The novelty of the method is that the PNCC features has not been suggested for SLI detection previously. The proposed method can be used for speech impairment detection and monitoring by therapist and doctors.