Detection of abrupt intensity changes in speech signals based on the receptive field concept

This paper explores the feasibility of automatic detection of vowel and word boundaries in speech signals represented in a sampled form of a neuromorphic model of the peripheral stage of the biological auditory perception system. An algorithmic approach is proposed that performs sequential processing of speech signals to detect the temporal boundaries of speech elements. At the initial stage, a sharpness intensity index is calculated using a system of temporal receptive fields, allowing for the identification of candidate segment boundaries. This is followed by event selection for the onset and offset of speech fragments and subsequent verification of vowel boundaries based on the local average spiking activity. To enhance the detection of word boundaries, the method further incorporates envelope analysis with adaptive thresholding, ensuring robustness and reproducibility of the segmentation results. The algorithm's performance was objectively evaluated using manually annotated time-aligned vowel boundaries, enabling the calculation of root mean square error and coincidence metrics. The results demonstrate high detection accuracy, reaching precision on the order of tens of milliseconds for word boundaries, thereby confirming the practical viability of the proposed approach. The method shows elevated sensitivity at vowel offsets, occasionally resulting in false detections, which indicates the need for incorporating adaptive contextual rules. The proposed system is suitable for real-time speech signal analysis and segmentation tasks.