This paper presents a word boundary detection technique based on frame classification using the nonlinear characteristics of speech. Bispectral analysis was used to classify speech frames into voiced, unvoiced and noise segments. To improve classification accuracy, bispectral features were combined with other features such as short time energy, zero-crossing rate, autocorrelation and high-to-low frequency ratio. Experimental results indicate that classification error decreases when bispectrum is combined with other features. Thus bispectral features can be used as supplementary to augment simple time domain features for demarcating word boundaries in speech. Validation of results was carried out by manual verification.