Elamur Research⌕
☰
home / papers / 2026 / Small Neural Networks as Models of Cross-Linguistic Speech Perception

Small Neural Networks as Models of Cross-Linguistic Speech Perception

Summary

The paper asks whether compact neural networks, rather than large speech-technology systems, can reproduce how a native language shapes the perception of non-native sounds. Its test case is Spanish-dominant adults hearing Catalan vowels. Earlier small-model work (Keidel, 2007) used only limited laboratory speech. Matusevych et al. (2023) merely suggested that supervised models might capture the Catalan /e/~/ɛ/ difficulty.

Here, four-layer feedforward networks learned to sort vowel tokens into the five Spanish categories, using data from the Spanish part of Multilingual Librispeech. They were then tested on vowels from a small Catalan read-speech corpus. All audio was force-aligned and preprocessed to approximate human auditory handling of vowels.

Because the networks can only output Spanish labels, Catalan /ɛ/ and /ɔ/ were expected to merge into native categories. They did: /ɛ/ was labelled /e/ or /a/, and /ɔ/ was labelled /o/. This matches the contrasts that Spanish-dominant listeners struggle with. Simulated discrimination, however, mostly failed to show the expected pattern.

The author argues that supervised learning matters more than model size. Small, interpretable networks are therefore viable cognitive models of the "L1 filter".

Key points

  • Data:
    • Spanish: Multilingual Librispeech, about 1000 h (918 h train, 10 h dev, 10 h test).
    • Catalan: the Catalan partition of Kjartansson et al. (2020), about 9.25 h.
  • Processing:
    • Alignment used the Montreal Forced Aligner, with Vargo's (2025) custom Catalan bundle for Catalan.
    • Loudness was normalised to -25 dB LUFS with pyloudnorm.
    • Each vowel slice had 480 time steps.
  • Main findings:
    • The models' errors fell on /e//ɛ/, /o//ɔ/ and /ɛ/~/a/, the contrasts that are hard for Spanish-dominant bilinguals.
    • Even networks with hidden layers of only 5 units reproduced this pattern.
    • The Spanish-52 model showed lower total variation distance on /e/~/ɛ/ than the baseline.
  • Limitations and open questions:
    • The baseline was likely under-trained.
    • Only one architecture was tested.
    • All models sometimes labelled /o/ as /u/, which is left for future work.
  • Resources: Code and forced alignments are public on GitHub (its-Annika/neural-speech-perception; its-Annika/es_ca_alignments).

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.scil-main.2

@inproceedings{shankwitz2026small,
  title = {{Small Neural Networks as Models of Cross-Linguistic Speech Perception}},
  author = {Shankwitz, Annika},
  booktitle = {Proceedings of the Society for Computation in Linguistics 2026},
  year = {2026},
  doi = {10.18653/v1/2026.scil-main.2},
  url = {https://doi.org/10.18653/v1/2026.scil-main.2}
}

References

22 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Frans Adriaans. 2024. Computational approaches to bilingual phonetics and phonology. In Mark Amengual, editor, Bilingual Phonetics and Phonology, pages 126–144. Cambridge University Press. find
  2. Diederik Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, San Diego, CA. find
  3. Mark Amengual. 2016. The perception and production of language-specific mid-vowel contrasts: Shifting the focus to the bilingual individual in early language input conditions. International Journal of Bilingualism, 20(2):133–152. find
  4. Oddur Kjartansson, Alexander Gutkin, Alena Butryna, Isin Demirsahin, and Clara Rivera. 2020. Open-source high quality speech datasets for Basque, Catalan and Galician. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages and Collaboration and Computi… find
  5. Catherine Best. 1995. A direct realist view of cross-language speech perception. In Winifred Strange, editor, Speech Perception and Linguistic Experience: Issues in Cross-Language Research, pages 171–204. Timonium, MD: York Press. find
  6. Catherine Best, Gerald McRoberts, and Elizabeth Goodell. 2001. Discrimination of non-native consonant contrasts varying in perceptual assimilation to the listener’s native phonological system. The Journal of the Acoustical Society of America, 109(2):775–794. find
  7. Tim Mahrt. 2016. PraatIO. find
  8. Yevgen Matusevych, Thomas Schatz, Herman Kamper, Naomi Feldman, and Sharon Goldwater. 2023. Infant phonetic learning as perceptual space learning: A crosslinguistic evaluation of computational models. Cognitive Science, 47(4):e13314. find
  9. Catherine Best and Michael Tyler. 2007. Nonnative and second-language speech perception: Commonalities and complementarities. In Ocke-Schwen Bohn and Michael McAuliffe, Michaela Socolof, Sarah Mihuc, Michael Wagner, and Morgan Sonderegger. 2017. Montreal forced aligner: Trainable text-speech alignme… find
  10. Baseline: 43.75%, Spanish-5 evaluated on Spanish: 27.5%, Spanish-52 evaluated on Spanish: 17.6%, Spanish-5 evaluated on Catalan: 48.75%, Spanish-52 evaluated on Catalan: 42.08%. find
  11. Michael McAuliffe and Morgan Sonderegger. 2022a. Spanish MFA acoustic model v2.0.0. Technical report. find
  12. Michael McAuliffe and Morgan Sonderegger. 2022b. Spanish (Spain) MFA G2P model v2.0.0. Technical report. find
  13. Christophe Pallier, Laura Bosch, and Núria Sebastián-Gallés. 1997. A limit on behavioral plasticity in speech perception. Cognition, 64(3):B9–B17. find
  14. Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. MLS: A large-scale multilingual dataset for speech research. In Interspeech 2020, pages 2757–2761. ISCA. find our card
  15. Joselyn Rodriguez, Kamala Sreepada, Ruolan Leslie Famularo, Sharon Goldwater, and Naomi Feldman. 2024. Self-supervised speech representations display some human-like cross-linguistic perceptual abilities. In Proceedings of the 28th Conference on Computational Natural Language Learning, pages 458–463… find
  16. Rebecca Ronquest. 2018. Vowels. In Kimberly Geeslin, editor, The Cambridge Handbook of Spanish Linguistics, pages 145–164. Cambridge University Press. find
  17. Thomas Schatz and Naomi Feldman. 2018. Neural network vs. HMM speech recognition systems as models of human cross-linguistic phonetic perception. In Conference on Cognitive Computational Neuroscience, Philadelphia, Pennsylvania. find
  18. Thomas Schatz, Naomi H. Feldman, Sharon Goldwater, Xuan-Nga Cao, and Emmanuel Dupoux. 2021. Early phonetic learning without phonetic categories: Insights from large-scale simulations on realistic input. Proceedings of the National Academy of Sciences, 118(7):e2001844118. find
  19. Thomas Schatz, Vijayaditya Peddinti, Francis Bach, Aren Jansen, Hynek Hermansky, and Emmanuel Dupoux. 2013. Evaluating speech features with the minimal-pair ABX task: analysis of the classical MFC/PLP pipeline. In Interspeech 2013, pages 1781– 1785, Lyon, France. find
  20. Núria Sebastián-Gallés and Salvador Soto-Faraco. 1999. Online processing of native and non-native phonemic contrasts in early bilinguals. Cognition, 72(2):111– 123. find
  21. Christian J. Steinmetz and Joshua D. Reiss. 2021. pyloudnorm: A simple yet flexible loudness meter in python. In 150th Audio Engineering Society Convention. find
  22. Julian Vargo. 2025. Catalan montreal forced alignment bundle. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shankwitz-2026-small-neural-networks-as-models/