IMPROVED TONE MODELING BY EXPLOITING ARTICULATORY FEATURES FOR MANDARIN SPEECH RECOGNITION

For the same tone pattern, different articulatory characteristics may make the pitch contour change. This paper applies articulatory features, which represent the articulatory information, as well as prosodic features to the tone modeling. Three kinds of tone models are trained to verify the effectiveness of articulatory features. Tone recognition experiments indicate significant improvement can be achieved when using both articulatory features and prosodic features. After the first pass search of a speech recognition system, tone models using new tonal features are employed to rescoring the N-best hypotheses, and a 6.5% relative reduction of character error rate is achieved.