Mandarin digit string speech recognition using linear discriminant analysis and tone discrimination model

Acoustic models based on the conventional hidden Markov model do not have high recognition performance for the connected Mandarin digit string, because highly confusable syllables exist. The state-specific linear discriminant analysis is adopted to reduce the substitution errors of confusable digits. The recognition rate for the isolated digit is increased from 97.16% to 99.32%; and the unknown length digit string from 86.5% to 88.18%. Furthermore experiments show that most of the typical confusions can be discriminated by the pitch contour patterns, the tone discrimination models are trained and the two-pass recognition algorithm to combine the acoustic model likelihood and the tone discrimination model likelihood is developed. By tone discrimination the relative digit string error rate is reduced by 37.4%. The unknown length digit string recognition rate and its digit recognition rate are increased from 88.18% and 97.54% to 92.6% and 98.21%, respectively.