Kernelized Deep Networks for Speech Signal Segmentation Using Clustering and Artificial Intelligence in Neural Networks

Speech signal segmentation is a key component of speech processing tasks, such as speaker recognition, speech recognition, and emotion recognition. Traditional approaches often have low accuracy and high computational complexity. To address this issue, we propose a novel approach using kernelized deep networks. The method utilizes Mel-Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coding (LPC) for feature extraction, with Spectral subtraction, wavelet denoising and Wiener filtering as preprocessing techniques. Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are used for dimensionality reduction. Spectral clustering using the graph Laplacian and normalized Laplacian matrix is then employed to cluster the speech signals into different segments. The Lanczos algorithm, Implicitly Restarted Arnoldi Method and Krylov subspace is used to train the networks and classify the signals. Results on TIMIT and BURN corpus demonstrate that the proposed approach outperforms traditional techniques in terms of segmentation accuracy and reduces computational complexity. This research highlights the potential of kernelized deep networks for speech segmentation and contributes to the field of speech processing.