Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data

doi:10.1109/TASLP.2021.3066274

CORC > 自动化研究所 > 中国科学院自动化研究所 > 模式识别国家重点实验室 > 智能交互

	Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data
	Bai, Ye 2; Yi, Jiangyan 1; Tao, Jianhua 1,3; Wen, Zhengqi 1; Tian, Zhengkun 2; Zhang, Shuai 2
刊名	IEEE-ACM TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING
	2021
卷号	29 页码:1340-1351
关键词	End-to-End language modeling speech recognition teacher-student learning transfer learning
ISSN号	2329-9290
DOI	10.1109/TASLP.2021.3066274
通讯作者	Yi, Jiangyan(jiangyan.yi@nlpr.ia.ac.cn) ; Tao, Jianhua(jhtao@nlpr.ia.ac.cn)
英文摘要	Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because of the end-to-end training, an AED model is usually trained with speech-text paired data. It is challenging to incorporate external text-only data into AED models. Another issue of the AED model is that it does not use the right context of a text token while predicting the token. To alleviate the above two issues, we propose a unified method called LST (Learn Spelling from Teachers) to integrate knowledge into an AED model from the external text-only data and leverage the whole context in a sentence. The method is divided into two stages. First, in the representation stage, a language model is trained on the text. It can be seen as that the knowledge in the text is compressed into the LM. Then, at the transferring stage, the knowledge is transferred to the AED model via teacher-student learning. To further use the whole context of the text sentence, we propose an LM called causal cloze completer (COR), which estimates the probability of a token, given both the left context and the right context of it. Therefore, with LST training, the AED model can leverage the whole context in the sentence. Different from fusion based methods, which use LM during decoding, the proposed method does not increase any extra complexity at the inference stage. We conduct experiments on two scales of public Chinese datasets AISHELL-1 and AISHELL-2. The experimental results demonstrate the effectiveness of leveraging external text-only data and the whole context in a sentence with our proposed method, compared with baseline hybrid systems and AED model based systems.
资助项目	National Key Research and Development Plan of China[2018YFB1005003] ; National Natural Science Foundation of China (NSFC)[61831022] ; National Natural Science Foundation of China (NSFC)[61901473] ; National Natural Science Foundation of China (NSFC)[61771472] ; National Natural Science Foundation of China (NSFC)[61773379] ; National Natural Science Foundation of China (NSFC)[173211KYSB20190049]
WOS关键词	NETWORK LANGUAGE MODELS
WOS研究方向	Acoustics ; Engineering
语种	英语
出版者	IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
WOS记录号	WOS:000640712800001
资助机构	National Key Research and Development Plan of China ; National Natural Science Foundation of China (NSFC)
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/44503]
专题	模式识别国家重点实验室_智能交互
通讯作者	Yi, Jiangyan; Tao, Jianhua
作者单位	1.Chinese Acad Sci, Inst Automat, NLPR, Beijing 100190, Peoples R China 2.Univ Chinese Acad Sci, Beijing 100190, Peoples R China 3.CAS Ctr Excellence Brain Sci & Intelligence Techn, Beijing 100190, Peoples R China
推荐引用方式 GB/T 7714	Bai, Ye,Yi, Jiangyan,Tao, Jianhua,et al. Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data[J]. IEEE-ACM TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING,2021,29:1340-1351.
APA	Bai, Ye,Yi, Jiangyan,Tao, Jianhua,Wen, Zhengqi,Tian, Zhengkun,&Zhang, Shuai.(2021).Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data.IEEE-ACM TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING,29,1340-1351.
MLA	Bai, Ye,et al."Integrating Knowledge Into End-to-End Speech Recognition From External Text-Only Data".IEEE-ACM TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCESSING 29(2021):1340-1351.