Elamur Research⌕
☰
home / papers / 2025 / Emergent morpho-phonological representations in self-supervised speech models

Emergent morpho-phonological representations in self-supervised speech models

Paper · 2025 · Jon Gauthier, Canaan Breiss, Matthew K Leonard et al.

Summary

The paper asks what linguistic structure a self-supervised speech model actually needs in order to tell spoken words apart, as opposed to structure it keeps for other purposes. The authors start from Wav2Vec2 Base, pretrained on LibriSpeech. For each of its layers they fit a linear "word probe" with a contrastive objective, which yields a subspace tuned for word discrimination. Word embeddings are averaged frames. Mikolov-style vector analogies then test how the English word-final sounds [z], [s] and [Iz] generalize across noun plurals, third-person verbs and monomorphemic words. Four possible sensitivities are examined: morphological, phonological, lexical and distributional.

In the word-optimized subspace these word pairs are related by a single global linear offset. This offset largely ignores whether the ending is a plural or a verb inflection, and also which allomorph appears. Instead it follows the distributional voicing/sibilance rule that governs word-final /-z/. The finding questions the assumption that recognition requires separate phonological and morphological levels of representation.

Key points

  • Base model: Wav2Vec2 Base, trained on 960 h of unlabeled LibriSpeech; it produces 768-dimensional frames covering 20 ms each, across 12 layers.
  • The word probe is a linear projection trained on LibriSpeech train-clean-100 (100 h, word-aligned). It uses a cosine hinge loss in which frames of the same word are positives and frames of other words are negatives.
  • The stimuli are 753 frequent nouns and 52 verbs that are unambiguous in part of speech.
  • Evaluation is the rank of the target word among 334,008 word tokens. The baseline uses a randomly chosen counterfactual word pair.
  • The regularity shows up from layer 1, peaks in intermediate layers and is visible in the principal components. Regression analysis shows that, compared with raw Wav2Vec, the probe has far weaker interaction effects of allomorph and inflection, with Worldlex frequencies included as controls.

Interesting ideas

  • I1. Comparing a general representation with a version optimized for one task shows which information the task really needs and which the model merely carries along.
  • I2. An analogy baseline built from a random counterfactual pair gives the "a:b::c:?" test a fair chance level instead of a bare score.
  • I3. The stimulus set includes words whose final sound is not an inflection at all, both rule-consistent (haze) and rule-inconsistent (fleece), so ending-based heuristics can be separated from true morphology.
  • I4. Regression on per-trial outcomes, with linguistic factors as interactions and frequency as a nuisance predictor, explains where a model fails rather than just averaging accuracy.
  • I5. The authors flag that a linear-offset assumption can produce false positives and false negatives, which is a reminder to check probing conclusions with another geometry or pooling scheme (they repeat the analysis with phoneme-level pooling).

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Probe gparse hidden states before and after task-specific training to see which morphological features attachment decisions actually rely on.
  • I2 (learn): Add such random-pair controls to evaluations of lemma and inflection relations as a negative control.
  • I3 (tokens): Build test lists of monomorphemic words ending in -s and -er so the en lemmatizer is checked against false stripping.
  • I4 (learn): Fit similar regressions on parser or tagger errors, with lemma frequency as a control, to find the linguistic conditions that actually cause mistakes.
  • I5 (world): Verify relations learned between verb classes or ontology entries with more than one similarity method before compiling them into the knowledge graph.

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

 Institutions

🎓University of Southern CaliforniaLos Angeles · United States162 works · 35,836 citations🎓University of California, San FranciscoSan Francisco · United States4 works · 9,908 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksDevelopmental and Educational Psychology · 499 works

Cite

Cite as Jon Gauthier, Canaan Breiss, Matthew K Leonard and Edward F. Chang. 2025. Emergent morpho-phonological representations in self-supervised speech models. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.emnlp-main.1425

@inproceedings{gauthier2025emergent,
  title = {{Emergent morpho-phonological representations in self-supervised speech models}},
  author = {Jon Gauthier, Canaan Breiss, Matthew K Leonard and Edward F. Chang},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1425},
  url = {https://doi.org/10.18653/v1/2025.emnlp-main.1425}
}

References

18 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449–12460.
  2. Kwanghee Choi, Ankita Pasad, Tomohiko Nakamura, Satoru Fukayama, Karen Livescu, and Shinji Watanabe. 2024. Self-supervised speech representations are more phonetic than semantic. In Proc. Interspeech 2024, pages 4578–4582. find
  3. Kwanghee Choi, Eunjung Yeo, Kalvin Chang, Shinji Watanabe, and David R Mortensen. 2025. Leveraging allophony in self-supervised speech models for atypical pronunciation assessment. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 2613–2628, Albuquerque, New Mexico. Association for Computational Linguistics. ↗
  4. Ewan Dunbar, Nicolas Hamilakis, and Emmanuel Dupoux. 2022. Self-supervised language learning from raw audio: Lessons from the zero resource speech challenge. IEEE Journal of Selected Topics in Signal Processing, 16(6):1211–1226. ↗
  5. David Embick, Ava Creemers, and Amy J Goodwin Davies. 2022. Morphology and the mental lexicon: Three questions about decomposition. ↗
  6. M Gareth Gaskell and William D Marslen-Wilson. 1997. Integrating form and meaning: A distributed model of speech perception. Language and cognitive Processes, 12(5-6):613–656. ↗
  7. Manuel Gimenes and Boris New. 2016. Worldlex: Twitter and blog word frequencies for 66 languages. Behavior research methods, 48:963–972. ↗
  8. Laura M Gonnerman, Mark S Seidenberg, and Elaine S Andersen. 2007. Graded semantic and phonological similarity effects in priming: evidence for a distributed connectionist approach to morphology. Journal of experimental psychology: General, 136(2):323. ↗
  9. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. In International Conference on Learning Representations. ↗
  10. Loren Lugosch. 2019. Librispeech alignments. find
  11. Kinan Martin, Jon Gauthier, Canaan Breiss, and Roger Levy. 2023. Probing self-supervised speech models for phonetic and phonemic information: a case study in aspiration. arXiv preprint arXiv:2306.06232. ↗
  12. James L McClelland and Jeffrey L Elman. 1986. The trace model of speech perception. Cognitive psychology, 18(1):1–86. ↗
  13. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26. ↗
  14. Dennis Norris. 1994. Shortlist: A connectionist model of continuous speech recognition. Cognition, 52(3):189–234. ↗
  15. Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206–5210. IEEE. find
  16. Gordon E. Peterson and Ilse Lehiste. 1960. Duration of syllable nuclei in english. Journal of the Acoustical Society of America, 32:693–703. ↗
  17. Ankita Pasad, Bowen Shi, and Karen Livescu. 2023b. Comparative layer-wise analysis of self-supervised speech models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. find
  18. Samuel A. Zimmerman and Stanley M. Sapon. 1958. Note on vowel duration seen cross-linguistically. Journal of the Acoustical Society of America, 30:152– 153. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gauthier-2025-emergent-morpho-phonological-representations-self/