There has been a tonne of study on use of ML for speech processing applications, particularly voice recognition, over the past few decades. This research proposed novel technique in speech signal processing and segmentation based on deep learning architectures. Here the input speech signal has been collected from the crime scene and this signal has been pre-processed using using K-means clustering (K-means C)for cluster the fragments of the input speech signal and process them for noise removal and signal artifacts removal. Here the segmentation is carried out for processed signal using Kernel based deep belief networks (KDBN). Experimental results demonstrate that proposed method outperforms the input speech signal based on both weighted accuracy (WA) and unweighted accuracy (UA).