Non-Autoregressive Chinese ASR Error Correction with Phonological Training

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5907 - 5917 July 10-15, 2022 ©2022 Association for Computational Linguistics Non-Autoregressive Chinese ASR Error Correction with Phonological Training Zheng Fang1,2, Ruiqing Zhang3∗, Zhongjun He3, Hua Wu3, Yanan Cao1,2 1Institute of Information Engineering, Chinese Academy of Sciences 2School of Cyber Security, University of Chinese Academy of Sciences 3Baidu Inc. No. 10, Shangdi 10th Street, Beijing, 100085, China 1,2{fangzheng, caoyanan}@iie.ac.cn 3{zhangruiqing01, hezhongjun, wu_hua}@baidu.com Abstract Automatic Speech Recognition (ASR) is an efficient and widely used input method that transcribes speech signals into text. As the er- rors introduced by ASR systems will impair the performance of downstream tasks, we intro- duce a post-processing error correction method, PhVEC, to correct errors in text space. For the errors in ASR result, existing works mainly focus on fixed-length corrections, modifying each wrong token to a correct one (one-to-one correction), but rarely consider the variable- length correction (one-to-many or many-to-one correction). In this paper, we propose an ef- ficient non-autoregressive (NAR) method for Chinese ASR error correction for both cases. Instead of conventionally predicting the sen- tence length in NAR methods, we propose a novel approach that uses phonological tokens to extend the source sentence for variable-length correction, enabling our model to generate pho- netically similar corrections. Experimental re- sults on datasets of different domains show that our method achieves significant improvement in word error rate reduction and speeds up the inference by 6.2 times compared with the au- toregressive model. 1 Introduction Errors introduced by automatic speech recognition (ASR) usually affect the performance of down- stream tasks such as phonetic search, speech trans- lation, etc. In recent years, ASR error correction techniques have been proposed (Anantaram et al., 2018; Mani et al., 2020; Zhao et al., 2021; Leng et al., 2021) to refine the ASR output and correct errors in text space. Without loss of generality, we study Chinese ASR error correction in this paper. Given the ASR result of an utterance, the goal of error correction is to generate a sentence with the wrongly recognized words corrected. Thus the ∗ This work was conducted at Baidu. Corresponding author: Ruiqing Zhang. ASR error correction can be modeled as a machine translation problem under conventional autoregres- sive sequence-to-sequence (Seq2Seq) framework (Guo et al., 2019; Hrinchuk et al., 2020; Mani et al., 2020). However, the autoregressive models suffer from inefficient decoding since the generation of each target token depends on previously generated characters (Figure 1 (a)). Furthermore, without considering the phonetic similarities, the method is prone to generate corrections with totally different pronunciation, as for the example, the error char- acter (“表”, biao) should be corrected into (“不 要”, bu yao) but the Seq2Seq model ignores phono- logical features and corrects it to a phonetically different correction (“不许”, bu xu). To speed up prediction, recent studies propose to take non-autoregressive (NAR) methods (Gu et al., 2018; Ren et al., 2020) for error correction, which generates target tokens …

Content truncated to 500 words