Federated, privacy-preserving multilingual speech models and European-language instruction tuning at UPC
3
papers here
3
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
2
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—
Output metrics are not yet sufficiently populated: this axis would measure total publication volume, venue types, and authorship roles over time. Only three 2026 preprints and a sparse OpenAlex record are currently available, so no score is assigned.
Impact—
Impact data is insufficient: this axis would measure citations, citing works, and influence on subsequent research. With an i10-index of 0 and no citation aggregates in the snapshot, scholarly impact cannot be assessed yet, so the axis is left unscored.
Focus0/100
The focus score of 0 reflects maximal topical breadth rather than inactivity: the three reviewed papers sit in three distinct fields, each at 0.333 of the field distribution — privacy-preserving speech ML, multilingual ASR, and multilingual instruction tuning — while OpenAlex topics spread across five areas led by NLP techniques (11), topic modeling (8), and speech recognition and synthesis (5). The common thread is multilingual modeling.
Collaboration32/100
Collaboration scores 32, consistent with a compact working circle: only 4 distinct coauthors appear in the OpenAlex record, with Jordi Luque (listed repeatedly), Fernando López and Carlos Escolano recurring. All three reviewed papers are multi-author arXiv preprints without a corresponding-author role, and authorship positions (first, middle, last) indicate participation across the author list within a small, stable team rather than a broad external network.
Activity75/100
Activity scores 75: the record shows 12 works in the last three years, with OpenAlex yearly counts rising from 1 (2023) to 6 (2024) and 6 (2026); no 2025 entry is recorded. All three Evolve-reviewed papers appeared in 2026, indicating a concentrated recent burst of preprint output across three research lines.
Reproducibility55/100
Reproducibility is scored 55, the mean across 3 reviewed papers, all rated medium readiness. Evolve notes concrete gaps: single-run federated benchmarks, a single training seed in EuroAlpaca with data unreleased, an untested gradient-inversion claim, and an unreported final privacy budget (ε) — recoverable with standard artifacts, but not yet demonstrated.
Research summary
Aleix Sant is a researcher affiliated with the Universitat Politècnica de Catalunya (UPC), Spain, whose recorded work centers on three interleaved lines: federated learning for multilingual Speech-LLMs, differential privacy for speech models, and multilingual instruction tuning for European languages. The OpenAlex-derived snapshot (h-index 2, i10-index 0) marks an early-career profile, but activity is high: 12 works in the last three years, with yearly counts of 1 (2023), 6 (2024) and 6 (2026). Topic frequencies are led by natural language processing techniques (11), topic modeling (8), speech recognition and synthesis (5), and privacy-preserving technologies (5). Evolve has reviewed three 2026 arXiv preprints: a first-authored EuroAlpaca study showing that task-preserving localisation of instruction data reverses MT-induced instruction-following loss in 50 languages; a last-authored component-aware differential privacy method for federated multilingual Speech-LLMs; and a middle-authorship benchmark of four Speech-LLMs under FedAvg/FedProx on MLS with 316 clients. Reproducibility scores 55 (mean across 3 papers), all rated medium readiness — reviewers flag single-run or single-seed results, an untested gradient-inversion claim, and unreleased data. Collaboration is concentrated: 4 distinct coauthors, with Jordi Luque, Fernando López and Carlos Escolano recurring. Output and impact axes are not yet scored due to insufficient citation data.
Collaboration pattern
Collaboration is compact: 4 distinct coauthors in the OpenAlex record, with recurring partners Jordi Luque, Fernando López and Carlos Escolano. Authorship on the three reviewed papers spans first, middle and last positions, all without corresponding-author status, suggesting a small stable team where roles rotate rather than a wide external network. The collaboration score of 32 reflects this narrow, repeatedly drawn co-author pool.
Research keywords
Highlights
EuroAlpaca: task-preserving localisation of instruction data (first author)
Component-aware differential privacy for federated Speech-LLMs (last author)
Benchmarking federated aggregation for multilingual Speech-LLMs (middle author)
Timeline
- 2023First recorded OpenAlex work appears, with a yearly count of 1.
- 2024Six works recorded in the OpenAlex yearly counts, a clear step up in productivity.
- 2026Speech line appears: middle-authorship federated benchmarking of four Speech-LLMs and last-authorship component-aware differential privacy, both arXiv preprints.
- 2026First-authored EuroAlpaca preprint on instruction-data localisation for European languages; the yearly count reaches 6 works.
Papers in the Evolve catalog
Frequent co-authors
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 scores and topic counts derive from OpenAlex only (identity: openalex); no ORCID is linked, so name resolution relies on public bibliographic data.
- Output and impact dimensions are unscored: citation aggregates are sparse (h-index 2, i10-index 0) and no venue-level data is available.
- All three reviewed papers are 2026 arXiv preprints; no peer-reviewed venue information is present in the data.
- No 2025 works appear in the yearly counts, and the co-author sample is small (4 distinct coauthors), so collaboration patterns may be under-sampled.