Fernando López

Research profile

Research summary

Fernando López is a researcher at Universidad Autónoma de Madrid (Spain) working at the intersection of speech processing and privacy-preserving machine learning. His Evolve-reviewed record centers on federated learning for multilingual speech large language models (Speech-LLMs). On the 2026 arXiv benchmark 'Federated Multilingual Speech-LLMs: Architecture and Aggregation Strategy Benchmarking' he is last author, evaluating four Speech-LLMs with FedAvg/FedProx and LoRA on MLS across 316 clients; Evolve found FedProx helps only EuroLLM and flagged single-run results. In 'Component-Aware Differential Privacy for Federated Multilingual Speech-LLMs' (middle author), an α-split two-pool per-layer clipping scheme addresses privacy-budget collapse, though Evolve notes the gradient-inversion claim is untested and the final ε unreported. OpenAlex topic data mirrors this dual profile: Speech Recognition and Synthesis (8), Speech and Audio Processing (5), Voice and Speech Disorders (4), Privacy-Preserving Technologies in Data (3), and Multimodal Machine Learning Applications (2), with the field split evenly (0.5/0.5) between multilingual-ASR machine learning and privacy-preserving machine learning. Metrics indicate an early-stage career: h-index 1, i10-index 0, 9 works in the recent three-year window (all recorded in 2026), and only 3 distinct co-authors (notably Aleix Sant and Jordi Luque), giving a low collaboration score of 28. Reproducibility readiness is rated medium (mean 55 across 2 papers).

Research keywords

Highlights

• Benchmarking federated aggregation for multilingual Speech-LLMs at 316-client scale

  As last author on this 2026 arXiv study, López contributed to a federated-learning benchmark of four Speech-LLMs on MLS across 316 clients, comparing FedAvg and FedProx with LoRA adaptation. Evolve's verdict: FedProx helps only EuroLLM, and per-component learning rates matter, though single-run results limit the strength of conclusions. Reproducibility readiness: medium.

• Component-aware differential privacy that prevents budget collapse in federated speech models

  In this middle-author 2026 paper, an α-split two-pool scheme for per-layer DP gradient clipping addresses privacy-budget collapse in federated multilingual Speech-LLMs. Evolve rates the diagnosis strong, while noting that the gradient-inversion claim remains untested and the final ε is unreported; reproducibility readiness is medium.

• One coherent program: private, federated multilingual speech-LLMs

  Read together, the two 2026 preprints sketch a single research program — first benchmarking federated architecture and aggregation choices for multilingual Speech-LLMs, then layering component-wise differential privacy on top of the same setting. Both involve the same core collaborators (Aleix Sant, Jordi Luque), positioning López as a contributing member of a compact research team rather than an independent lead.

Timeline

1. 2026Published on arXiv 'Federated Multilingual Speech-LLMs: Architecture and Aggregation Strategy Benchmarking' — a federated-learning benchmark of four Speech-LLMs on MLS across 316 clients; López is last author.
2. 2026Published on arXiv 'Component-Aware Differential Privacy for Federated Multilingual Speech-LLMs' — an α-split two-pool per-layer clipping scheme against privacy-budget collapse; López is a middle author.
3. 2026OpenAlex snapshot: 9 works recorded in 2026 (the full recent three-year window), h-index 1, i10-index 0, and 3 distinct co-authors.

Papers in the Evolve catalog

• 2026-09-20

  Federated Multilingual Speech-LLMs: Architecture and Aggregation Strategy Benchmarking

• 2026-09-10

  Component-Aware Differential Privacy for Federated Multilingual Speech-LLMs

Frequent co-authors