Confusable Chinese speech recognition based on HMM/SVM two-level architecture

The recognition rate for confusable speech is still low in state-of-the-art Chinese speech recognition systems based on HMM. The inherent defects of HMM are analyzed, then a two-level-architecture recognition framework combining HMM and SVM is proposed. A confidence estimation module is adopted to improve the performance and efficiency of the system. The information obtained by Viterbi decoding is utilized to construct new classes of feature for SVM, which solves the problem that the conventional SVM cannot directly process variable length sequences. The relevant issues, such as confidence estimation, classification feature extraction and SVM recognizer construction, are addressed. The experimental results of confusable Chinese speech show that compared with the hybrid HMM/SVM based system the proposed method can highly improve the recognition rate with little impact on the running speed.