Abstract：As the current pruning thresholds can not take decoding speed and accuracy into account at the same time in continuous speech recognition, a joint optimization algorithm of multi-dimension pruning thresholds parameters is proposed. The pruning thresholds, including the main beam pruning, the word end pruning, the number of active modes and the tokens, are mainly studied in the proposed algorithm. The multi-objectives theory is adopted to optimize these parameters jointly. And then the strategy of segment-based dynamic thresholds pruning is introduced to deal with the results. The experimental results show that the performance of decoder is improved, the search space of decoding gets effective control, and the request of speed and accuracy can be satisfied.
[1] Huang Xuedong,Acero A,Hon H W.Spoken Language Processing-A Guide Theory Algorithm and System Development.Upper Saddle River,USA: Prentice Hall,2001 [2] Jang J S R,LIN S S.Optimization of Viterbi Beam Search in Speech Recognition // Proc of the International Symposium on Chinese Spoken Language Processing.Taibei,China,2002: 114-117 [3] Wendt S,Fink G A,Kummert F.Dynamic Search-Space Pruning for Time-Constrained Speech Recognition // Proc of the 7th International Conference on Spoken Language Processing.Denver,USA,2002: 377-380 [4] Xie Lingyu.Research on Fast Algorithm of Real-Time Speech Recognition System.Ph.D Dissertation.Beijing,China: Institute of Acoustics of Chinese Academy of Sciences,2004 (in Chinese) (谢凌云.实时语音识别系统的快速算法研究.博士学位论文.北京:中国科学院声学研究所,2004) [5] Liu Yin.Research and Implementation of Large Vocabulary Continuous Speech Recognition System.Ph.D Dissertation.Beijing,China: Tsinghua University,2005 (in Chinese) (刘 盈.大词表连续语音识别系统的研究与实现.博士学位论文.北京:清华大学,2005) [6] Huang Pin,Meng Yonggang.Optimization Theory and Methods.Beijing,China: Tsinghua University Press,2009 (in Chinese) (黄 平,孟永钢.最优化理论与方法.北京:清华大学出版社,2009) [7] EI Hannani A,Hain T.Automatic Optimization of Speech Decoder Parameters.IEEE Signal Processing Letters,2010,17(1): 95-97 [8] Mak B,Ko T.Min-Max Discriminative Training of Decoding Parameters Using Iterative Linear Programming // Proc of the Interspeech.Brisbane,Australia,2008: 915-918
2012, Vol. 25 
