Automatic speech segmentation is a key step for building large vocabulary and continuous speech recognition systems. An alternative method is to use manual speech segmentation, which is tedious, time consuming, subjective and error prone. Many automatic speech segmentation methods have been proposed based on linguistic information such as phonetic transcription (phonetic string) but for real time systems this phonetic transcription not always be available. In This paper we propose a new algorithm for Arabic speech Consonant and Vowel (C/V) segmentation without linguistic information. This new method is based on wavelet transform and spectral tilt and focuses on searching the transient between Consonant and Vowel parts in certain levels from wavelet packet decomposition. To verify the proposed scheme, some experiments have been performed using set of words, each word recorded six times. The accuracy rate is about 88.3% for Consonant/vowel segmentation. This rate remains fixed with low SNR value as well as high SNR.

Keywords:

Automatic speech recognition (ASR), Modern Standard Arabic (MSA),

Wavelet Packet decomposition, Spectral Tilt.