计算机工程与应用 ›› 2011, Vol. 47 ›› Issue (32): 163-166.

• 数据库、信号与信息处理 • 上一篇    下一篇

多分类器区分性组合在二次解码中的应用

黄 浩,李兵虎   

  1. 新疆大学 信息科学与工程学院 多语种信息技术实验室,乌鲁木齐 830046
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2011-11-11 发布日期:2011-11-11

Discriminative combination of multiple local classifiers in lattice rescoring

HUANG Hao,LI Binghu   

  1. Laboratory of Multi-lingual Information Technology,Department of Information Science and Engineering,Xinjiang University,Urumqi 830046,China
  • Received:1900-01-01 Revised:1900-01-01 Online:2011-11-11 Published:2011-11-11

摘要: 提出利用基于隐马尔可夫模型的谱特征模型、基于高斯混合模型的声调分类器以及基于多层感知器的音素分类器模型的组合来提高语音识别中二次解码中的识别率。在模型组合中,使用上下文相关的模型权重加权模型得分,并使用区分性训练来优化上下文相关权重来进一步改进识别结果。对人工选取各种上下文相关权重集合进行了性能评估,连续语音识别实验表明,使用局部分类器进行二次解码能够明显降低系统误识率。在模型组合中,使用当前音节类型及左上下文相结合的模型权重集合能够最大程度降低系统误识率。实验表明该方法得到的识别结果优于基于谱特征与基频特征和音素后验概率特征合并得到特征组合的识别系统。

关键词: 区分性模型组合, 语音识别, 多层感知器, 区分性训练

Abstract: The combination of the hidden Markov model based spectral acoustic model,multi-layer perceptron based phoneme classifier and Gaussian mixture model based tone classifier in lattice rescoring is proposed.Moreover,discriminative model weight training is applied to tune the impact of the heterogeneous models according to different phonetic contexts for better model interpolation.Experimental results on continuous speech recognition show significant improvement can be obtained using the combination of the models.Four context dependent weighting schemes for discriminative trained scaling factors are evaluated.It is also shown introducing left contexts can obtain the best recognition accuracy.Results have also shown tree based model combination is superior to the system based on feature space combination.

Key words: discriminative model combination, speech recognition, multi-layer perceptron, discriminative training