Sentence segmentation for speech processing

Abstract

Automatic sentence segmentation of speech is a process of identifying the end of a sentence. It is used for improving the output after speech recognition and helps in making the recognition output more readable. It is generally a two-class problem which involves the identification of a boundary characterizing the sentence part and the non-sentence part. An Automatic Speech Recognition (ASR) system receives a stream of speech signal and produces a un-annotated stream of words. The regions of silences in the inter word boundaries of the output of ASR are detected. The prosodic features of the input speech signal that are given to the ASR are extracted for a particular duration of time which includes pause, rhyme, slope, minimum, maximum and mean features. Once the features are extracted, SVM classifier uses all the features and discriminates each word boundary as sentence or non-sentence boundary.

Non-Stationarity-Based Adaptive Segmentation Applied to Voice Disorder Discrimination

“…Signal segmentation involves breaking down a continuous speech signal into smaller, more manageable frames. The following are the goals of segmentation: to maintain the speech as a wide-sense stationary (WSS) process for several feature extraction approaches [20], optimize feature extraction and machine learning procedures [27], and identify and isolate the individual speech sounds or words in a continuous stream of speech [28]. Stationarity implies that the statistical properties of the signal remain constant over time.…”