Automatic Segmentation of Continuous Speech on Word and Phrase Level based on Suprasegmental Features

This article investigates whether it is possible to segment continuous speech on word and phrasal level by examination of suprasegmental parameters, in case of bound stress languages like Hungarian and Finnish. The final aim is to increase the robustness of speech recognition on language modelling level by the detection of word and phrase boundaries and so we can significantly decrease the searching space during the decoding process. In Hungarian language if stress is present, it is always on the first syllable of the word stressed. Thus if stressed syllables can be detected, word boundaries are also detectable with reliable accuracy. We have developed different algorithms based either on rule based or on data-driven approach. Rule based algorithms and HMM based method are compared. The best results were obtained by data-driven algorithms using the time series of fundamental frequency and energy together. Syllable length was found much less effective hence it was discarded. By using of suprasegmental features word boundaries can be marked by a good accuracy, if we allow not to find all of them. The method we evaluated is easily adaptable to other fixed stress languages. To investigate this we adapted our datadriven method to Finnish language and got similar results.