Speech Stammer Detection by Spectral Features Based Artificial Neural Network

In many applications of computer-assisted speech analysis, the detection of stammer from speech signals is essential. For training the stammer detection model, a strong feature is required to distinguish the speech signal as a stammer or non-stammer. The proposed work employs Mel Frequency Cepstral Coefficients (MFCC), Delta Delta MFCC (D2MFCC), Pitch, Spectral Flux, and Spectral Centroid to extract the dominant features from speech. These features are utilized to train a Multilayer Perceptron (MLP) Neural Network (NN) using a Bayesian Regularization (BR). When compare to NN training, classifiers, and the state of the art, this technique gives superior accuracy and training. Experiments have also been conducted to confirm the primacy of fused spectral features over other standard features used in literature and it has been demonstrated that the proposed features when combined with MLP-BR provide better accuracy to the level of 99.2% over the collected speech dataset.