A Novel Chinese Mandarin Speech Indexing Method Based on Confusion Network Using Tone Information

We proposed a novel method for Chinese Mandarin speech indexing on the basis of confusion network in this paper. Chinese Mandarin is a tonal language which has 5 basic tones including neutral tone. The method utilized this characteristic of Chinese to model tone. Firstly, we obtained a tonal syllable confusion network (CN) through a conversion after an introduction of traditional lattice. Subsequently, we used MSD-HMM model to get a tone model, with the use of which we merge the CN into an acoustic model. Finally we got our best speech indexing system with a precision of 84.6%. The experiment results showed that the merged confusion network system had a better performance comparing with the former system with an improvement of 4.2%. And the run speed of CN can be nearly 5 times faster than the traditional lattice.