{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/minimum-word-error-rate-training-for","title":"Minimum Word Error Rate Training for Attention-based Sequence-to-Sequence Models","arxiv_id":"1712.01818","date":"2017-12-05","proceeding":null,"authors":["Rohit Prabhavalkar","Tara N. Sainath","Yonghui Wu","Patrick Nguyen","Zhifeng Chen","Chung-Cheng Chiu","Anjuli Kannan"],"abstract":"Sequence-to-sequence models, such as attention-based models in automatic\nspeech recognition (ASR), are typically trained to optimize the cross-entropy\ncriterion which corresponds to improving the log-likelihood of the data.\nHowever, system performance is usually measured in terms of word error rate\n(WER), not log-likelihood. Traditional ASR systems benefit from discriminative\nsequence training which optimizes criteria such as the state-level minimum\nBayes risk (sMBR) which are more closely related to WER. In the present work,\nwe explore techniques to train attention-based models to directly minimize\nexpected word error rate. We consider two loss functions which approximate the\nexpected number of word errors: either by sampling from the model, or by using\nN-best lists of decoded hypotheses, which we find to be more effective than the\nsampling-based method. In experimental evaluations, we find that the proposed\ntraining procedure improves performance by up to 8.2% relative to the baseline\nsystem. This allows us to train grapheme-based, uni-directional attention-based\nmodels which match the performance of a traditional, state-of-the-art,\ndiscriminative sequence-trained system on a mobile voice-search task.","url_abs":"http://arxiv.org/abs/1712.01818v1","url_pdf":"http://arxiv.org/pdf/1712.01818v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"minimum-word-error-rate-training-for","repo_url":"https://github.com/TeaPoly/CTC-MWER","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"unanswered"}},{"paper_slug":"minimum-word-error-rate-training-for","repo_url":"https://github.com/TeaPoly/CTC-OptimizedLoss","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"automatic-speech-recognition-2","task_name":"Automatic Speech Recognition"},{"task_slug":"automatic-speech-recognition","task_name":"Automatic Speech Recognition (ASR)"},{"task_slug":"speech-recognition","task_name":"Speech Recognition"},{"task_slug":"speech-recognition-1","task_name":"speech-recognition"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1712.01818","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}