基于语言模型的调度语音智能识别研究

Research on Scheduling Speech Intelligent Recognition Based on Language Model

针对调度语音识别过程中单遍解码词图生成算法所生成词图精度较差的问题,研究基于语言模型的调度语音智能识别方法。构建由训练过程和识别过程组成的调度语音智能识别模型,训练过程中该模型提取语音数据的语音向量序列构建声学子模型,利用语言子模型训练文本数据构建语音词图,识别过程中对声学子模型、语音词图以及发音词典实施语音解码与搜索获取最优词序列,基于最优词序列完成调度语音智能识别。测试结果显示研究方法所生成的词图精度较高,可准确识别调度语音。

Aiming at the problem of poor accuracy of word graph generated by single pass decoding word graph generation algorithm in the process of scheduling speech recognition,a scheduling speech intelligent recognition method based on language model is studied.A scheduling speech intelligent recognition model composed of training process and recognition process is constructed.In the training process,the model extracts the speech vector sequence of speech data to construct the phonological sub model,and uses the language sub model to train the text data to construct the speech word map.In the recognition process,the phonological sub model,speech word map and pronunciation dictionary are decoded and searched to obtain the optimal word sequence.Scheduling speech intelligent recognition is completed based on the optimal word sequence.The test results show that the word graph generated by the research method has high accuracy and can accurately recognize the scheduling speech.