Word Boundary Estimation For Continuous Speech Using Higher Order Statistical Features

Detection of the start and the end time of words in a continuous speech plays a crucial role in enhancing the accuracy of Automatic Speech Recognition (ASR). Hence, addressing the problem of efficiently demarcating word boundaries is of prime importance. In this paper, we introduce two new acoustic features based on higher order statistics called Density of Voicing (DoV) and Variability of Voicing (VoV) obtained from the bispectral distribution, which when used along with the popular prosodic cues helps in drastically reducing the recognition error rate involved. An ensemble of three different models has been designed to minimize the false alarms, during word boundary detection, by maximizing the uncorrelatedness in prediction from each model. Finally, the majority-voting rule was used to decide if the given frame encompasses a word boundary. The contribution of the work lies in: (i) Converting word boundary detection into a supervised learning problem (ii) Introduction of two new higher order statistical features (iii) Using ensemble methods to find the best model for prediction. Experimental results for NTIMIT Database shows the efficacy of the proposed method and its robustness to noise added during telephonic transmission.