Research on Segment Acoustic Model Based Mandarin LVCSR

  • Authors:

  • Wenju Liu;Yun Tang;Shouye Peng

  • Affiliations:

  • National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China 100190;National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China 100190;National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China 100190

  • Venue:
  • ISNN 2009 Proceedings of the 6th International Symposium on Neural Networks: Advances in Neural Networks - Part II
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

SM has shown a better performance than HMM in connected word recognition system; however, no reports we have read show that SM has been applied in LVCSR as decoding acoustic model because of the restriction of its complexity. We have preliminarily built a SM based mandarin LVCSR system which adopts CART and global tying to tie the parameters in the triphone models and the fast SM algorithm, CF algorithm and two-level pruning to enhance the speed of decoding. The system achieves 87.09% syllable accuracy in Test-863 data corpus within 4 real times. We believe SM offers an alternative choice for LVCSR system though further research for its fast algorithms by rational utilization of its structure information.