Automatic speech segmentation using neural network and phonetic transcription

A new algorithm for automatic segmentation of speech based on its phonetic transcription is proposed. The specific features of this algorithm are a new iterative self-learning procedure to find the temporal alignment between feature vectors and phonetic transcription; no preassumptions about statistical speech properties or phonetical rules; and no required pretraining. The general structure of the segmentation system is shown. The core of the segmentation procedure is an iterative loop consisting of a neural phoneme classifier, a time-alignment algorithm and the retraining of the neural classifier. The segmentation of the sentence `nine two seven eight nine ten´ is given