Discriminative incorporation of explicitly trained tone models into lattice based rescoring for Mandarin speech recognition

Explicit tone modeling has been widely discussed in recent Mandarin speech recognition research. In this paper, a discriminative method of incorporating explicitly trained tone models into lattice based rescoring is proposed. The method is to use discriminative trained model weights to scale the acoustic model and tone model distributions. The weights are trained by the Minimum Phone Error using the Extended Baum Welch algorithm. To take into account different phonetic contexts, various model weighting schemes are evaluated. A smoothing technique is introduced to make model weight training more robust to over fitting. The proposed method is evaluated on tonal syllable output speech recognition tasks on a Mandarin LVCSR database. Results show the proposed method has achieved significant error reduction than traditional global weight approach. Comparison with the traditional embedded tone modeling is also made, which shows the importance of the proposed method when explicit tone modeling approach is applied.