A study on robust segmentation and location of tone nuclei in Chinese continuous speech

Tone nuclei in continuous speech are regarded as efficient targets for either tone recognition or intonation function decomposition. The paper presents our statistically robust method to segment and locate tone nuclei in continuous speech. The method includes: an iterative segmental K-means segmentation of the tonal F0 contours, which is further aided with t-test based segment amalgamation; a linear discriminant function based tone nucleus discriminator, whose features are selected by the sequential feature selection method. The developed system achieved 97.5% correct tone nuclei on a speaker dependent task. The tone recognizer based on the detected tone nuclei improved tone recognition rate by over 6% more than the baseline ones using the full tonal syllable features.