Abstract

Speech recognition system requires segmentation of speech waveform into fundamental acoustic units. Segmentation is a process of decomposing the speech signal into smaller units. Speech segmentation could be done using wavelet, fuzzy methods, Artificial Neural Networks and Hidden Markov Model. Speech segmentation is a process of breaking continuous stream of sound into some basic units like words, phonemes or syllable that could be recognized. Segmentation could be used to distinguish different types of audio signals from large amount of audio data, often referred as audio classification. The speech segmentation can be divided into two categories based on whether the algorithm uses previous knowledge of data to process the speech. The categories are blind segmentation and aided segmentation.The major issues with the connected speech recognition algorithms were the vocabulary size will be larger with variation in the combination of words in the connected speech and the complexity of the algorithm is more to find the best match for the given test pattern. To overcome these issues, the connected speech has to be segmented into words using the attributes of speech. A methodology using the temporal feature Short Term Energy was proposed and compared with an existing algorithm called Dynamic Thresholding segmentation algorithm which uses spectrogram image of the connected speech for segmentation.

Authors

A. Akila¹, E. Chandra²
D.J. Academy for Managerial Excellence, India¹, Bharathiar University, India²

Keywords

Short Term Energy, Missed Detection Percentage, Deviation Percentage, Dynamic Thresholding Segmentation, Temporal Feature based Segmentation