A robust C/V segmentation algorithm for Cantonese

A real time robust consonant/vowel (C/V) segmentation algorithm for Cantonese is proposed. Segmentation is begun by finding the pitch information, zero-crossing rate, energy profile and linear prediction coefficients in a stable vowel region. In each comparison, the speech signals within one pitch period are used/or calculated to avoid incorrect location of a repeating speech unit due to pitch change. The pitch period is found by center-clipping of speech signals and maximum point searching. The search for the C/V segmentation is performed by comparison of pitch information, zero-crossing rate, energy profile and linear prediction coefficients by searching towards the consonant. With this technique, the false location of the CV segmentation on the C/V transient region can be avoided. Also, with an accuracy of 95.1% and a low computation complexity, this algorithm is suitable for automated C/V segmentation and speech recognition