Applying Batch Normalization to Hybrid NN-HMM Model For Speech Recognition

作者：Zhan, Hongjian^*; Chen, Guilin; Lu, Yue

来源：7th Chinese Conference on Pattern Recognition (CCPR), China,Sichuan,Chengdu, 2016-11-05 to 2016-11-07.

DOI：10.1007/978-981-10-3005-5_35

• Acoustic model
• Covariate shifts
• HMM models
• Network modeling
• Original model
• Speaker-independent speech recognition
• Training procedures
• Word error rate

Batch Normalization has showed success in image classification and other image processing areas by reducing internal covariate shift in deep network model's training procedure. In this paper, we propose to apply batch normalization to speech recognition within the hybrid NN-HMM model. We evaluate the performance of this new method in the acoustic model of the hybrid system with a speaker-independent speech recognition task using some Chinese datasets. Compared to the former best model we used in the Chinese datasets, it shows that with batch normalization we can reach lower word error rate (WER) of 8%-13% relatively, meanwhile we just need 60% iterations of original model to finish the training procedure.

• 出版日期2016
• 单位华东师范大学