Abstract: In order to lighten the heavy computational burden of one-pass lattice-generating algorithms for speech recognition, a fast two-pass decoding algorithm was proposed on the basis of the forward-backward language model. The forward and backward language models were applied to the first and second decoding processes separately. Furthermore, some optimization rules were given to reduce the impact of language model mismatch and to avoid its side-effects on recognition results. The experimental results show that this algorithm quickens the decoding process without decreasing the recognition accurate rate.

Key words: speech recognition, lattice, language model, tow-pass decoding, backward extension

摘要： 为了克服语音识别中单遍解码词图生成算法速度较慢的缺点，提出一种基于前后向语言模型的两遍快速解码算法。两遍解码分别采用前向与后向语言模型,同时通过优化以减少前后向语言模型不匹配对识别结果造成的影响。实验证明，该算法在保持识别准确率的基础上有效地提升了解码速度。

关键词: 语音识别, 词图, 语言模型, 两遍解码, 后向扩展