Lstm-based speech segmentation trained on different foreign languages

Abstract

This paper describes experiments on speech segmentation by using bidirectional LSTM neural networks. The networks were trained on various languages (English, German, Russian and Czech), segmentation experiments were performed on 4 Czech professional voices. To be able to use various combinations of foreign languages, we defined a reduced phonetic alphabet based on IPA notation. It consists of 26 phones, all included in all languages. To increase the segmentation accuracy, we applied an iterative procedure based on detection of improperly segmented data and retraining of the network. Experiments confirmed the convergence of the procedure. A comparison with a reference HMM-based segmentation with additional manual corrections was performed.