Fuzzy Logic Inference System for Voiced-Unvoiced-Silence Classification of Malaysian English Isolated Words based on Time-domain Features

The ability of a system to automatically detect speech signals such as, voiced-unvoiced-silence can have a great impact to the accuracy and computation time in isolated- word speech recognition systems. The presence of a long inter-syllable silence in Malaysian English speech is due to the habit of spelling and reading methods in syllabic manner of the national Malay language at school. Further, manual segmentation is very tedious, laborious and error- prone for large vocabularies and sample size. Time- domain features such as short-time energy and zero- crossing rate are proven to be very efficient and simple computations. Since frame-based speech analysis does not consider the start and end of a particular phoneme, the best and easiest way to make decision for the important features should be inferred from human expert knowledge. In this paper, fuzzy inference system is build from membership functions and fuzzy rules derived from a simple statistical analysis of the speech data from males and females and three ethnic groups of Malaysian English speakers.