基于改进Conformer的新闻领域端到端语音识别

End-to-End Speech Recognition in News Field Based on Conformer

目前,开源的中文语音识别数据集大多面向通用领域,缺少面向新闻领域的开源语音识别语料库,因此该文构建了面向新闻领域的中文语音识别数据集CH_NEWS_ASR,并使用ESPNET-0.9.6框架的RNN、Transformer和Conformer等模型对数据集的有效性进行了验证,实验表明,该文所构建的语料在最好的模型上CER为4.8%,SER为39.4%。由于新闻联播主持人说话语速相对较快,该文构建的数据集文本平均长度为28个字符,是Aishell_1数据集文本平均长度的2倍;且以往的研究中训练目标函数通常为基于字或词水平,缺乏明确的句子水平关系,因此该文提出了一个句子层级的一致性模块,与Conformer模型结合,直接减少源语音和目标文本的表示差异,在开源的Aishell_1数据集上其CER降低0.4%,SER降低2%;在CH_NEWS_ASR数据集上其CER降低0.9%,SER降低3%,实验结果表明,该方法在不增加模型参数量的前提下能有效提升语音识别的质量。

The open source Chinese speech recognition data sets are usually developed for the general domain.This paper constructs a news-oriented Chinese speech recognition data set named CH_NEWS_ASR,and verifies the validity of the data set by the RNN,Transformer and Conformer models under ESPNET-0.9.6 framework.As news broadcasters speak relatively fast,the average text length in this dataset is 28 characters,which is 2 times of the average text length of Aishell_1 dataset.In this paper,we propose a sentence-level consistency module combined with the Conformer model to directly reduce the representation differences between source speech and target text.Experiments demonstrate that,on the Aishell_1 dataset,the CER is reduced by 0.4%and the SER by 2%;on the CH_NEWS_ASR dataset,the CER is reduced by 0.9%and the SER by 3%.