Elamur Research⌕
☰
home / papers / 2026 / Intent vs. Surface: Recovering Acoustic Realization from Modern ASR for Pronunciation Training

Intent vs. Surface: Recovering Acoustic Realization from Modern ASR for Pronunciation Training

Paper · 2026 · Seongjin Park

Summary

Computer-assisted pronunciation training usually compares an ASR transcript with the prompt the learner read. Modern recognizers, however, rely on language-model priors. They tend to output the word the speaker meant rather than the sounds actually produced, which hides learner errors. The author calls this "intent bias" and measures it across eight ASR systems from three architecture families on two L2 English corpora, L2-ARCTIC and speechocean762. Systems with lower WER turn out to conceal more mispronunciations.

As a remedy, the paper proposes surface-faithful reranking. An LM-free phoneme recognizer supplies a reference phone string, and each N-best hypothesis is converted to phonemes with g2p-en. The candidate closest under a PanPhon articulatory-feature edit distance is then used for word-level error detection, while the ordinary top-1 transcript is still shown to the learner. The method needs no retraining and no access to model internals. It cuts false acceptance of mispronunciations by about six points, with consistent gains across age groups and L1s. The work shows that WER alone is a poor criterion for choosing ASR in educational tools.

Key points

  • Overcorrection is systematic. Across autoregressive, hybrid CTC-RNNT and CTC-only systems, lower WER correlates with higher overcorrection. With Parakeet, 82.0% of speechocean762 utterances contain at least one masked mispronunciation in the top hypothesis.
  • Setup: the main systems are NVIDIA Parakeet (FastConformer, hybrid CTC-RNNT) and Whisper large-v2 via faster-whisper, each with K=10 hypotheses. Phoneme distance uses 24-dimensional PanPhon feature vectors with Hamming-based substitution costs, and no weights are tuned.
  • False acceptance rate (FAR) on severe errors drops from 80.8% to 74.8% on L2-ARCTIC (six L1s, 24 speakers). On the speechocean762 test split (n=2,497), FAR drops from 70.6% to 65.0%. The cost is a false rejection rate rise of 1.7–3.0pp, mostly on short function words.
  • The gains are capped by the content of the N-best list. The K=10 oracle FAR is still 56.9%, and widening Whisper's beam to K=20 adds only 1.4pp. The phoneme recognizer itself has a 19.2% error rate on L2-ARCTIC.
  • To complement WER, the author introduces a new metric called Pronunciation Preservation Rate, defined as PPR = 1 − FAR. On the speechocean762 test split it ranges from 25.8% to 59.0%. Future work includes pronunciation-aware decoding and multi-pronunciation lattices.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 worksSignal Processing · 493 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 519 works

Cite

Cite as Seongjin Park. 2026. Intent vs. Surface: Recovering Acoustic Realization from Modern ASR for Pronunciation Training. Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bea-1.23

@inproceedings{park2026intent,
  title = {{Intent vs. Surface: Recovering Acoustic Realization from Modern ASR for Pronunciation Training}},
  author = {Seongjin Park},
  booktitle = {Proceedings of the 21st Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.bea-1.23},
  url = {https://doi.org/10.18653/v1/2026.bea-1.23}
}

References

23 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: a framework for self-supervised learning of speech representations. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook, NY, USA. Curran Associates Inc.
  2. David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin. 2016. Pan-Phon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages… find
  3. The COLING 2016 Organizing Committee. find
  4. Ambra Neri, Catia Cucchiarini, Helmer Strik, and Lou Boves. 2002. The pedagogy-technology interface in computer assisted pronunciation training. Computer Assisted Language Learning, 15(5):441–467. find
  5. Kyubyong Park and Jongseok Kim. 2019. g2pE. https://github.com/Kyubyong/g2p. ↗
  6. Enrique Cámara-Arenas, Cristian Tejedor-García, Carlos J. Tomás-Vázquez, and David Escudero-Mancebo. 2023. Automatic pronunciation assessment vs. automatic speech recognition: A study of conflicting conditions for L2-English. Language Learning & Technology, 27(1):1–19. find
  7. Linkai Peng, Kaiqi Fu, Binghuai Lin, Dengfeng Ke, and Jinsong Zhan. 2021. A Study on Fine-Tuning wav2vec2.0 Model for the Task of Mispronunciation Detection and Diagnosis. In Interspeech 2021, pages 4448–4452. find
  8. Chen Chen, Ruizhe Li, Yuchen Hu, Sabato Marco Siniscalchi, Pin-Yu Chen, EngSiong Chng, and Chao-Han Huck Yang. 2024. It’s never too late: Fusing acoustic information into large language models for automatic speech recognition. In The Twelfth International Conference on Learning Representations. find
  9. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 28492–2…
  10. Maxine Eskenazi. 2009. An overview of spoken language technology for education. Speech Communication, 51(10):832–844. Spoken Language Technology for Education. find
  11. Yuan Gong, Ziyi Chen, Iek-Heng Chu, Peng Chang, and James Glass. 2022. Transformer-based multi-aspect multi-granularity non-native English speaker pronunciation assessment. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7262–7266.
  12. Daniel Korzekwa, Jaime Lorenzo-Trueba, Szymon Za-porowski, Shira Calamaro, Thomas Drugman, and Bozena Kostek. 2021. Mispronunciation detection in non-native (l2) English with uncertainty modeling. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)…
  13. Wai-Kim Leung, Xunying Liu, and Helen Meng. 2019. CNN-RNN-CTC based end-to-end mispronunciation detection and diagnosis. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8132–8136. find
  14. Kun Li, Xiaojun Qian, and Helen Meng. 2017. Mispronunciation detection and diagnosis in l2 English speech using multidistribution deep neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 25(1):193–207. find
  15. Somshubra Majumdar and Nithin Rao Koluguri. 2024. Pushing the boundaries of speech recognition with NVIDIA NeMo parakeet ASR models. NVIDIA Technical Blog. find
  16. Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699–2712, Online. Association for Computational Linguistics.
  17. SYSTRAN. 2023. faster-whisper. https://github.com/SYSTRAN/faster-whisper. ↗
  18. Silke M. Witt and Steve J. Young. 2000. Phone-level pronunciation scoring and assessment for interactive language learning. Speech Communication, 30(2— 3):95–108. find
  19. Yongkook Won. 2025. Assessing the efficacy of word error rate as a proxy for pronunciation quality. Journal of Second Language Pronunciation, 11(3):394–422. find
  20. Xiaoshuo Xu, Yueteng Kang, Songjun Cao, Binghuai Lin, and Long Ma. 2021. Explore wav2vec 2.0 for Mispronunciation Detection. In Interspeech 2021, pages 4428–4432. find
  21. Klaus Zechner, Derrick Higgins, Xiaoming Xi, and David M. Williamson. 2009. Automatic scoring of non-native spontaneous speech in tests of spoken English. Speech Communication, 51(10):883–895. find
  22. Junbo Zhang, Zhiwen Zhang, Yongqing Wang, Zhiyong Yan, Qiong Song, Yukai Huang, Ke Li, Daniel Povey, and Yujun Wang. 2021. speechocean762: An Open-Source Non-Native English Speech Corpus for Pronunciation Assessment. In Interspeech 2021, pages 3710–3714.
  23. Guanlong Zhao, Sinem Sonsaat, Alif Silpachai, Ivana Lucic, Evgeny Chukharev-Hudilainen, John Levis, and Ricardo Gutierrez-Osuna. 2018. L2-ARCTIC: A Non-native English Speech Corpus. In Interspeech 2018, pages 2783–2787. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/park-2026-intent-vs-surface-recovering-acoustic/