Scalable language model look-ahead for LVCSR

In this paper a new computation and approximation scheme for Language Model Look-Ahead (LMLA)is introduced. Themain benefit of LMLA is sharper pruning of the search space during the LVCSR decoding process. However LMLA comes with its own cost and is known to scale badly with both LM n-gram order and LM size. The proposed method tackles this problem with a divide and conquer approach which enables faster computation without additional WER cost. The obtained results allowed our system to participate in the real-time task of the ESTER Broadcast News transcription evaluation campaign for French.

[1]  Hermann Ney,et al.  Language-model look-ahead for large vocabulary speech recognition , 1996, Proceeding of Fourth International Conference on Spoken Language Processing. ICSLP '96.

[2]  Holger Schwenk,et al.  USING NEURAL NETWORK LANGUAGE MODELS FOR LVCSR , 2004 .

[3]  Mehryar Mohri,et al.  A weight pushing algorithm for large vocabulary speech recognition , 2001, INTERSPEECH.

[4]  Carmen García-Mateo,et al.  Fast LM look-ahead for large vocabulary continuous speech recognition using perfect hashing , 2002, 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing.

[5]  Georges Linarès,et al.  Phoneme Lattice Based A* Search Algorithm for Speech Recognition , 2002, TSD.

[6]  Xavier L. Aubert,et al.  An overview of decoding techniques for large vocabulary continuous speech recognition , 2002, Comput. Speech Lang..

[7]  Hermann Ney,et al.  A comparison of two LVR search optimization techniques , 2002, INTERSPEECH.

[8]  Hermann Ney,et al.  Improved lexical tree search for large vocabulary speech recognition , 1998, Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP '98 (Cat. No.98CH36181).