Catherine Arnett

University of Cambridge

Unclaimed · AI-structured profileOpenAlex

University of Cambridge researcher co-authoring bilingual LM suites and 102-model ASR releases with top reproducibility

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

The output axis is not yet computable for this profile: the current snapshot lacks the publication counts and venue records needed to score it. Once filled in, it would measure total and per-year research output across the catalogued venues.

Impact—

No impact data is available yet, so this axis cannot be scored. In a complete profile it would summarize citation counts, h-index-style aggregates, and downstream uptake of the reviewed papers.

Focus0/100

Focus scores 0 out of 100, reflecting an even 50/50 field distribution between speech processing / NLP and computational psycholinguistics / NLP. The two catalogued papers fall one into each field, so current published work spans two adjacent areas rather than concentrating in a single dominant one.

Collaboration39/100

Collaboration scores 39 out of 100 with 6 distinct coauthors recorded. Arnett appears as a middle-list, non-corresponding author on both catalogued papers, and the frequent-coauthor list includes Shivam Singh, Aditya Yadavalli, Alex Warstadt, Suchir Salhan, James A. Michaelov, and Paula Buttery, suggesting team-based rather than solo work.

Activity—

Activity data is insufficient in this snapshot, so no score is given. Once available, the axis would track publication cadence and recent-year engagement derived from paper timestamps.

Reproducibility100/100

Reproducibility is the strongest axis: the two reviewed papers average a score of 100 (reproMean 100 across 2 reproPapers), and both carry high reproducibility readiness. Beetle ships checkpoints for 285 bilingual and 45 monolingual models, and BuzzASR releases 102 fine-tuned models, consistent with a resource-release working style.

Research summary

Catherine Arnett is a researcher at the University of Cambridge working across computational psycholinguistics and speech / natural language processing. The Evolve catalog holds two 2026 arXiv papers with Arnett as a middle-list, non-corresponding author. Beetle releases a bilingual model suite — 285 bilingual and 45 monolingual language models with checkpoints — and finds staged curricula best predict second-language reading times, tempered by single-seed training and second-language-share confounds. BuzzASR ships 102 monolingual Whisper fine-tunes that beat Whisper-large-v3 on 77 of 102 languages, roughly a 2.8× cut in mean character error rate, with train/test domain match tempering the SOTA claim. Bibliometrically, the field distribution splits evenly, 50/50 across the two fields, giving a focus score of 0; collaboration scores 39 with 6 distinct coauthors, including Shivam Singh, Aditya Yadavalli, Alex Warstadt, Suchir Salhan, James A. Michaelov, and Paula Buttery. Reproducibility is the standout axis: the two reviewed papers average a score of 100 with high reproducibility readiness. Output, impact, and activity axes are missing from the current snapshot, so this profile reads best as an early-stage, resource-release-centered record.

Collaboration pattern

Arnett's recorded collaboration footprint is compact: 6 distinct coauthors, a collaboration score of 39/100, and middle-list (non-corresponding) authorship on both catalogued papers. The frequent-coauthor list — Shivam Singh, Aditya Yadavalli, Alex Warstadt, Suchir Salhan, James A. Michaelov, and Paula Buttery — points to a stable team anchored at the University of Cambridge producing large shared model releases rather than many loose one-off partnerships.

Research keywords

bilingual language modelssecond-language acquisitioncurriculum learningcomputational psycholinguisticsreading time predictionsurprisalautomatic speech recognitionWhisper fine-tuninglow-resource languagestokenizer replacementmultilingual ASRFLEURS

Highlights

Bilingual LM suite scaled to 285 bilingual and 45 monolingual models
The Beetle suite releases language models with checkpoints across 285 bilingual and 45 monolingual variants, making second-language processing modelling directly reproducible; Evolve rates its reproducibility readiness as high.
Staged curricula best predict L2 reading times
Across the Beetle suite, staged curricula emerge as the best predictors of second-language reading times, though single-seed training and second-language-share confounds temper the strength of the claim.
102 monolingual ASR models beat Whisper-large-v3 on 77 languages
BuzzASR fine-tunes Whisper into 102 monolingual models that outperform Whisper-large-v3 on 77 of 102 languages, cutting mean character error rate by about 2.8×; train/test domain match tempers the SOTA claim, and reproducibility readiness is rated high.

Timeline

2026BuzzASR: 102 monolingual Whisper fine-tunes released on arXiv, with strong coverage of low-resource languages.
2026Beetle: bilingual model suite (285 bilingual + 45 monolingual LMs with checkpoints) released on arXiv.
2026Both catalogued papers are Evolve-reviewed arXiv preprints with middle-list authorship and high reproducibility readiness.

Papers in the Evolve catalog

Frequent co-authors

Shivam Singh×1 Aditya Yadavalli×1 Alex Warstadt×1 Suchir Salhan×1 James A. Michaelov×1 Paula Buttery×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric dimension scores are derived from OpenAlex-style public bibliographic records only; no ORCID is linked, so name resolution may be imperfect for a relatively common name.
The catalog contains only 2 Evolve-reviewed papers (both 2026 arXiv preprints); peer-review status is not captured in the current data.
Output, impact, and activity axes are marked missing — no publication counts, citation aggregates, or cadence data were available to score them.
No country field was provided; institutional affiliation is listed only as the University of Cambridge.