语种确认中基于段长的语言模型修正方法

Revision of Language Model Based on Duration in Language Verification

为了改善语言模型提出了一种新型的语言模型修正方法,先利用音素段长的筛选性来获得部分音素的语言模型修正因子,再对修正因子加以权重用于纠正语言模型的偏差,从而能够有效提高系统识别性能。在CALLFR IEND Corpus上进行的测试中,系统单个前端的最好性能达到等错率下降11.54%,整个系统等错率下降了6.93%,最后的等错率为9.50%。

A novel approach is proposed to improve estimation of language model probability, revision factors and weights are used for achieving new language model estimation. Evaluation on CALLFRIEND Corpus shows 11.54% best relative improvement in EER on single phone recognition system and 6. 93% relative improvement in EER on parallel phone recognition system, the final EER is 9. 50%.