基于动态贝叶斯网络的语音识别及音素切分研究

Research on DBN-based continuous speech recognition and phoneme segment

研究了一种基于动态贝叶斯网络(dynamic bayesian networks,DBN)的语音识别建模方法,利用GMTK(graphical model tool kits)工具构建音素级音频流DBN语音训练和识别模型,同时与传统的基于隐马尔可夫的语音识别结果进行比较,并给出词与音素的切分结果。实验表明,在各种信噪比测试条件下,基于DBN的语音识别结果与基于HMM的语音识别结果相当,并表现出一定的抗噪性,音素的切分结果也比较准确。

This paper described a dynamic Bayesian network （DBN） based technique on continuous speech recognition. The word recognition accuracies and phoneme segment accuracies of the DBN based system （ implemented using the graphical model tool kit） were compared with those from classical HMM. Results show that under various SNRs, DBN based system and HMM based system has similarity performance for speech recognition and phoneme segment, especially in much lower SNR circumstance, DBN get even much better performance than HMM.