Speech segmentation is a foundational task involving the division of a speech signal into discrete acoustic units, such as phonemes or syllables. Despite its importance, accurately identifying boundaries between these units remains challenging due to the inherent variability in speech patterns. Our study addresses the segmentation of speech into phonemes without requiring transcript data or any prior knowledge of these linguistic units. The method utilized the extraction of maxima from the first derivative of Mel-frequency cepstral coefficients (MFCCs), known as \(\Delta \) -MFCCs, and applied an adaptive threshold to identify phoneme boundaries. These features were chosen due to their effectiveness in capturing dynamic changes in the speech signal that correspond to significant transitions. Unlike static features, \(\Delta \) -MFCCs are particularly suited for representing time-based variations, enabling the detection of subtle shifts in the speech signal. The true strength of this approach lies in the use of an adaptive thresholding mechanism. Unlike a fixed threshold, the adaptive threshold was calibrated to adjust dynamically based on the intensity of the detected peaks in the \(\Delta \) -MFCCs. The generated boundaries were validated using a matching technique applied to reference utterances. To assess the efficacy of the proposed method, a comprehensive set of experiments was conducted using both Arabic and English speech datasets. Our approach achieved notable performance, demonstrating an F1-score of 96% and 85% in Arabic and English, respectively. Furthermore, it was compared against state-of-the-art segmentation techniques, showcasing its superiority in accurately segmenting speech signals. The success of our approach underscores its potential for enhancing speech recognition systems, contributing to the development of highly accurate systems capable of processing diverse speech patterns.
