Abstract: Compared with traditional speech recognition system, airplane conjunction speech recognition system has background noise, and requires a high recognition rate and so on. According to these features, this paper proposes a EMD-SDC method with empirical mode decomposition and shifted delta cepstral features. Empirical mode decomposition with characteristics of AM FM can substantially increase endpoint detection accuracy under complex airplane noise environment. Shifted delta cepstral which is composed of first-order differential spectral of the speech frames, can capture the time sequence information depending on the structure of the language well. This method is tested for airplane traffic collision avoidance system database, experimental result shows that the airplane conjunction speech recognition system with EMD-SDC method can overcome cabin background noise and achieve a higher recognition rate in the low SNR.

Key words: empirical mode decomposition, shifted delta cepstral, airplane conjunction speech recognition

摘要： 机载连接词语音识别系统与传统语音识别系统相比，具有背景噪声大，系统识别率要求高等特点。依据这些特点，提出了一种基于经验模态分解增强和位移差分倒谱特征的EMD-SDC连接词语音识别方法。经验模态分解的调频调幅特性，可以有效提高机载复杂噪声背景下的端点检测准确度，位移差分倒谱特征由语音帧的一阶差分谱连接扩展而成，能够更好地提取依赖于语言结构的时序信息。该方法对机载交通预警避撞系统提示语音库进行测试，实验结果表明，采用EMD-SDC方法的机载连接词语音识别系统，能够很好地克服机舱背景噪声干扰，在低信噪比条件下实现较高的识别率。

关键词: 经验模态分解, 位移差分倒谱, 机载连接词语音识别