For a dynamic network based large vocabulary continuous speech recognizer, this paper proposes a fast language model (LM) look-ahead method using extended iV-gram model. The extended N-gram model unifies the representations and score computations of the LM and the LM look-ahead tree, and thus greatly simplifies the decoder implementation and improves the LM look-ahead speed significantly, which makes higher-order LM look-ahead possible. The extended N-gram model is generated off-line before decoding starts. The generation procedure makes use- of sparse-ness of backing-off N-gram models for efficient look-ahead score computation, and uses word-end node pushing and score quantitation to compact the model's storage space. Experiments showed that with the same character error rate, the proposed method speeded up the overall recognition speed by a factor of 5 ~ 9 than the traditional dynamic programming method which computes LM look-ahead scores on-line during the decoding process, and that using higher-order LM look-ahead algorithm can achieve a faster decoding speed and better accuracy than using the lower-order look-ahead ones.%针对基于动态解码网络的大词汇量连续语音识别器,本文提出了一种采用扩展N元文法模型进行快速语言模型(Language model,LM)预测的方法.扩展N元文法模型统一了语言模型和语言模型预测树的表示与分数计算方法,从而大大简化了解码器的实现,极大地提升了语言模型预测的速度,使得高阶语言模型预测成为可能.扩展N元文法模型在解码之前离线生成,生成过程利用了N元文法的稀疏性加速计算过程,并采用了词尾节点前推和分数量化的方法压缩模型存储空间大小.实验表明,相比于采用动态规划在解码过程中实时计算语言模型预测分数的传统方法、本文提出的方法在相同的字错误率下使得整个识别系统识别速率提升了5～9倍,并且采用高阶语言模型预测可获得比低阶预测更优的解码速度与精度.
展开▼
