发音特征在维汉语音识别中的应用

Application of Articulatory Feature in Uygur and Mandar in Speech Recognition

采用传统谱特征作为输入进行语音识别通常会受到声学环境差异的影响。为此,提出汉语和维语音素和音位的对应规则,并将这种规则应用于基于发音特征的语音识别系统。训练神经网络多层感知器,获取语音信号各类发音特征的后验概率,将其与美尔频率倒谱系数(MFCC)拼接后送入隐马尔科夫模型进行声学模型训练。将不同发音特征分别与传统MFCC特征进行组合并给出测试结果。实验结果表明,当汉语声带状况和送气发音特征与传统MFCC组合时,以及维语的发音方式和声带状况特征与MFCC组合之后,系统误识率较低。

Speech recognition based on traditional spectral feature is liable to be influenced by the acoustic conditions of the environment.Articulatory Feature（AF） is robust to such conditions.In this paper,the rules of phonetic mapping to AF of Mandarin and Uighur speech are derived.The neural networks are trained to obtain posterior probability of AF.The features are combined with Mel Frequency Cepstral Coefficient（MFCC） and are used to train the hidden Markov based acoustic model.Experimental results show that by combining the MFCC with the feature of voicing or aspiration in Mandarin,the feature of voicing or manner in Uighur,significant error reductions can be obtained.