农产品信息采集作业场景下目标用户的语音识别研究

Research on the Automatic Speech Recognition of Target Users in the Scene of Agricultural Information Collection

农产品信息采集场所通常以大型集贸批发市场为主,其语音场景又以单语言、多口音、多目标、高噪声为特点,在此特定领域开展语音识别应用研究的关键在于解决多语音数据集之间的声学、通道和环境的不匹配问题。本文考虑在小词汇量任务下,使用标准普通话和两种宁夏地方口音的数据,从目标用户群体中采集语音数据,在单语言多口音环境下分别训练连续概率密度隐马尔可夫模型(CDHMM)和子空间高斯混合模型(SGMM),通过基线实验对比了两种声学建模方法的性能特点。

Agricultural information collection usually occurs in large wholesale markets with the speech scene of single language,multiple accents,multiple targets and strong noise.The key of application research on automatic speech recognition in this scene is to solve the mismatching problems of acoustics,channel and environment among multiple speech data sets. Considering the tasking scene with small vocabulary,this paper collected the speech data from target users based on the standard mandarin and two local dialects in Ningxia, and trained the CDHMM model and SGMM model respectively in the environment of single language and multiple accents,and compared the features of this two acoustic modeling methods through baseline test.