A phone segmentation method and its evaluation on Mandarin speech corpus

Abstract

This paper presents a phone segmentation method without a prior knowledge about the text contents. The proposed method is an unsupervised phone boundary detection based on band-energy tracing technique. It demonstrates a better performance than those previous works when the method was applied to TIMIT corpus. But the performance degrades when the method is applied to a Mandarin Chinese speech database, TCC300 corpus. The evaluation on this Mandarin speech corpus reveals some interesting facts that may cause the difficulty in detecting phone boundaries. We have proposed some ideas that may be helpful in future study for improving the phone segmentation method.