A dynamic term discovery strategy for automatic speech recognizers with evolving dictionaries

作者：

Highlights：

• ASR OOV processing strategy that allows a dynamic vocabulary adaptation.

• An unsupervised adaptation of the LM towards the current speech is produced.

• A model and a set of modules are proposed for the dynamic arrangements.

• Strategy evaluated on two long test corpora and multiple vocabulary sizes.

• Positive tendency in WER, with significant improvements in most of the cases.

摘要

•ASR OOV processing strategy that allows a dynamic vocabulary adaptation.•An unsupervised adaptation of the LM towards the current speech is produced.•A model and a set of modules are proposed for the dynamic arrangements.•Strategy evaluated on two long test corpora and multiple vocabulary sizes.•Positive tendency in WER, with significant improvements in most of the cases.

论文关键词：Dynamic term discovery,Out-of-vocabulary words,Language model adaptation,Automatic speech recognition

论文评审过程：Received 2 October 2019, Revised 25 January 2021, Accepted 4 March 2021, Available online 13 March 2021, Version of Record 2 April 2021.

论文官网地址：https://doi.org/10.1016/j.eswa.2021.114860