A hybrid unsupervised segmentation algorithm for arabic speech using feature fusion and a genetic algorithm (July 2018)

Speech is the most natural form of human communication. Major achievements have been made in developing systems that automatically recognize human speech and respond to it. An important preprocessing step in speech recognition systems (and other applications) is segmentation. Such a step is important in identifying the high-level semantics of speech sounds, including syllables, consonants, vowels..., and so on. There are basically two general approaches used for speech segmentation, namely, explicit and implicit (so-called with and without linguistic reference). Explicit segmentation uses a bottom-up process based on the concept of fixed-size frames. Such a framework is usually used on conjunction with hidden Markov models. The varying frame size or sample-by-sample approaches are mainly used in implicit segmentation techniques, which are based on the detection of spectral distortions. The main objective of this paper is to develop a novel speech segmentation algorithm for Arabic language. In this application, we focus on the accurate segmentation of Quran recitation. The proposed system starts with a set of initial segmentations using three basic speech features: Entropy, zero crossings, and energy. The segmentation results obtained are then fused at the output level using a genetic algorithm-based optimization scheme. Together with the segmentation results, we also introduce the concept of speech units to model the fundamental entities representing Quran recitation. Our results show an enhanced performance in segmentation of about 20% over that of traditional single-feature-based segmentation techniques.