MASR中文语音识别 MASR是一个基于端到端的深度神经网络的中文普通话语音识别项目。 原理 MASR使用的是门控卷积神经网络（Gated Convolutional Network），网络结构在Facebook在2016年提出的Wav2letter。但是使用的激活函数不是ReLU HardTanh ，而不是GLU （门控线性单元）。因此根据我的实验，使用GLU的收敛速度比HardTanh要快。如果您想要研究卷积网络用于语音识别的效果，这个项目可以作为一个参考。 以下用字错误率CER来假定模型的表现，CER =编辑距离/句子长度，越低越好 大致可以理解为1-CER就是识别准确率。 模型使用AISHELL-1数据集训练，共150小时的录音，覆盖了4000多个汉字。工业界使用的语音识别系统通常使用至少10倍于本项目的录音数据来训练，同时使用特定场景的语料来训练语言模型，所以，不要期待本项目可以

The Lancaster Corpus of Mandarin Chinese (LCMC) is designed as a Chinese match for the FLOB and FROWN corpora for modern British and American English. The corpus is suitable for use in both monolingual research into modern Mandarin Chinese and cross-linguistic contrast of Chinese and British/American English. The corpus sampled 15 written text categories including news, literary texts, academic prose and official documents etc published in P. R. China in the earlier 1990s for a total of approximately 1 million words. The same sampling frame and period as FLOB/FROWN were used in LCMC. The corpus is marked up for text categories, sample file numbers, paragraphs, sentences and tokens. Linguistic annotations undertaken on the corpus include tokenization and part-of-speech tagging. The whole corpus is annotated at the word level and includes orthographic and morphological annotations. The tagging system used was produced by the Institute of Computing Science Chinese Lexical Analysis System (ICTCLAS), the Chinese Academy of Sciences. The corpus is encoded in Unicode (UTF-8) and marked up in XML. The corpus comes with a User Manual detailing corpus design specifications and part-of-speech tags. The XML structure of the corpus was validated using the parser built in Xaira. Part-of-speech tagging of all aspect markers was manually checked.