New benchmarks and models for low-resource languages, medical and brain foundation models, AI in space, and innovation kits

In this issue:

• Paza: New ASR benchmarks and models for low-resource languages, grounded in real community use
• Vibe Kit & Innovation Kits: Turning research ideas into working prototypes faster
• MIRA: A medical time-series foundation model for zero-shot forecasting on real-world health data
• OrbitalBrain: Distributed machine learning directly onboard satellite constellations
• Brain foundation models: Interpretable cognitive load estimation for next-generation BCIs

NEW BENCHMARK AND MODELS 

Paza: Introducing automatic speech recognition benchmarks and models for low resource languages 

Microsoft is releasing Paza, an automatic speech recognition (ASR) platform, including PazaBench, the first ASR leaderboard for low-resource, underrepresented languages. Built for and tested in the communities served, Paza launches with 39 African languages and 52 human-centered state-of-the-art models, evaluated on three key metrics across leading public and community datasets. These models were fine-tuned using minimal data and tested with farmers on everyday mobile devices, focusing on six Kenyan languages: Swahili, Dholuo, Kalenjin, Kikuyu, Maasai, and Somali. 

Building on Project Gecko, which aims to expand AI benefits to the global majority, Paza addresses the failure of speech systems in low-resource environments, where many languages remain unrecognized and non-Western accents are misunderstood. Community testers assess Paza models on real devices in real contexts. Upcoming playbooks will provide practical guidance on dataset creation, fine-tuning with minimal data, and evaluation, empowering researchers to develop systems grounded in real human use. 

NEW TOOLS 

Vibe Kit and Innovation Kits help turn research into results  

Vibe Kit is a customized agentic coding environment for rapid prototyping and integration of Microsoft Research innovations using GitHub Copilot in VS Code. Innovation Kits live inside Vibe Kit, offering research project-specific knowledge, instructions, context, and code samples. Practitioners can get started quickly, adapt reusable patterns, and integrate capabilities with fewer false starts. This accelerates the journey from idea to impact: quickly test whether a research innovation fits your use case, build demos that show what's possible, and turn successful experiments into real breakthroughs. To get started, install an Innovation Kit, open the quick start, and let GitHub Copilot guide you through build, test, and iteration—prototypes in hours, not weeks. 

NEW TOOL 

MIRA: Medical Time Series Foundation Model for Real-World Health Data 

MIRA is a foundation model designed to learn a unified representation space across heterogeneous clinical datasets and support zero-shot forecasting in real-world healthcare settings. Unlike conventional time-series models that operate on fixed sampling rates or task-specific feature spaces, MIRA is built to natively handle irregular and clinically diverse signals. By combining continuous-time encoding, frequency-aware specialization, and neural dynamics modeling, MIRA generalizes robustly across environments and variables. 

MIRA is pretrained on 454 billion time points from large-scale clinical corpora spanning both intensive care unit (ICU) physiological signals and hospital electronic health record (EHR) time-series. This covers a rich range of sampling frequencies, including minute-level vitals, hourly labs, waveform segments, and multi-day clinical indicators, allowing MIRA to serve as a unified backbone capable of strong out-of-distribution generalization. In extensive evaluations, MIRA achieves state-of-the-art (SOTA) zero-shot forecasting performance across diverse clinical benchmarks—demonstrating strong robustness under dataset shift, irregular sampling, and multimodal temporal variations. 

Key features 

• Continuous-Time Rotary Positional Encoding (CT-RoPE)—provides a principled way to embed irregular timestamps while preserving temporal geometry, enabling robust reasoning across arbitrary sampling patterns. 

• Frequency-specialized Mixture-of-Experts—allows different experts to specialize on physiological signs, improving transfer across diverse clinical signals. 

• Neural ordinary differential equation ODE extrapolation—models latent dynamics continuously over time, enabling forecasting at arbitrary future timestamps. 

NEW RESEARCH 

OrbitalBrain: A Distributed Framework For Training ML Models in Space 

The paper introduces OrbitalBrain, a distributed machine learning (ML) framework designed to optimize ML training on Earth observation (EO) nanosatellites. These satellites face significant challenges, particularly limited downlink bandwidth that results in delays of days or weeks in transmitting high-resolution images back to Earth for training. OrbitalBrain overcomes this by utilizing the satellites' onboard compute resources, bandwidth, and power more effectively, enabling in-space ML training without relying on ground stations. 

OrbitalBrain balances data transfer, model aggregation, and local training through a predictive scheduler that optimizes satellite resources based on orbital, power, and storage forecasts. This approach is both energy-efficient and scalable, facilitating real-time or near-real-time model training and updates directly within satellite constellations. 

In performance evaluations with two satellite constellations and multiple ML tasks, OrbitalBrain achieves a 1.52× to 12.4× speedup in time-to-accuracy compared to state-of-the-art ground-based or federated learning approaches, while also achieving higher model accuracy. The system also adapts effectively to varying conditions, such as cloud obstructions and different image resolutions. 

OrbitalBrain represents a significant advancement in satellite-based AI, enabling distributed intelligence across satellite constellations, enhancing satellite-based applications like forest fire and flood detection. 

NEW RESEARCH 

Cognitive Load Estimation Using Brain Foundation Models and Interpretability for BCIs 

Brain foundation models (BFMs) are large, pre-trained neural networks designed to extract generalizable representations from neural signals, such as EEG (electroencephalography) data. This study investigates how BFMs can support continuous cognitive load monitoring, tackling longstanding challenges related to scalability, cross-subject variability, and interpretability in EEG analysis.  

The authors develop a scalable pipeline for long-term cognitive-load estimation using BFM-derived features and introduce a group-average channel alignment method that improves generalization across heterogeneous EEG sensor layouts. They further adapt Partition SHAP for neural time-series data, enabling interpretable, neuroscience-consistent estimates of feature importance. A longitudinal evaluation across multiple days tracks how cognitive load and neural markers evolve over repeated learning sessions. 

The results show that as participants become more familiar with a learning task, cognitive load decreases, while the relevance of frontal and prefrontal brain activity increases. These brain regions are associated with cognitive control and decision-making, suggesting that BFMs preserve meaningful neural signatures of engagement even as the user’s task performance varies. At the same time, BFM features improve cognitive load estimation accuracy compared with prior EEG-based approaches.  

Overall, the study positions BFMs as an effective and interpretable tool for real-time cognitive load monitoring. Their ability to provide extended-context inference, adapt across individuals, and maintain neuroscientific relevance highlights their potential for next generation BCIs that respond to user engagement and support personalized learning in non-invasive, cost-effective settings suitable for everyday use.