摘要
普通话声调识别参数除常用的基音轮廓外,基音的一阶差分、能量及能量的一阶差分等也具一定的声调特征。实验结果表明:如果将各种参数同时作为一个BP模型的输入参数,声调识别率不但没有提高,反而显著下降,因此,该文提出了将各种参数分别训练一个各自的BP网络,再将这些网络的输出作为另一高层BP网络的输入的普通话声调识别方法。另外,针对上声的特点提出了一种改进的基音平滑算法。这些方法的运用使系统的声调识别率达到90.05%。
In a mandarin tone recognition system,the parameter is usually pitch contour.But the tone can also be partly characterized by pitch difference,energy and energy difference.The experimental result shows that if all these parameters are input ted to one neural network,the recognition results are not any better but much lower than the system only using pitch contour.This paper describes a new method of using those parameters.It is that every kind of parameter is used to train one different BP model,and all the outputs of these BP models as the inputs of the higher level BP model.An improved pitch smooth method is put forward by the characterization of the third tone.The experimental result shows that using these new methods,an accuracy rate of90.05%can be achieved in mandarin tone recognition.
出处
《计算机工程与应用》
CSCD
北大核心
2004年第25期96-99,共4页
Computer Engineering and Applications
关键词
声调识别
语音识别
神经网络
tone recognition,speech recognition,neural network
