Automated radiology report generation is expanding across imaging modalities and anatomical regions, yet its evaluation remains misaligned with diagnostic priorities. Existing metrics either rely on surface-level text similarity or require slow, costly generative inference. This limits their ability to detect clinically significant errors, such as hallucinated or omitted abnormal findings, and prevents their use as efficient reward signals during training. We introduce POLAR (POlarity-Aware Long-context Anatomy-conditioned Regression), an encoder-based evaluation metric built on ModernBERT. POLAR combines long-context inputs for multi-organ reports, anatomy conditioning for targeted body-region evaluation, and polarity-aware regression that explicitly distinguishes normal from abnormal findings, yielding a more clinically meaningful assessment than prior metrics. Across real-world, radiologist-annotated CXR benchmarks and expert preference datasets, POLAR achieves competitive alignment with radiologist judgments while running 11x faster than generative evaluators, making it practical for large-scale benchmarking and as a training-time reward signal. We publicly release POLAR and the POLAR-Synth dataset, along with a fully open-source, privacy-preserving pipeline for data generation, training, and inference, enabling the metric to be extended to new modalities and anatomies.

Current state of the art acoustic models can easily comprise more than 100 million parameters. This growing complexity demands larger training datasets to maintain a decent generalization of the final decision function. An ideal dataset is not necessarily large in size, but large with respect to the amount of unique speakers, utilized hardware and varying recording conditions. This enables a machine learning model to explore as much of the domain-specific input space as possible during parameter estimation. This work introduces Common Phone, a gender-balanced, multilingual corpus recorded from more than 76.000 contributors via Mozilla’s Common Voice project. It comprises around 116 hours of speech enriched with automatically generated phonetic segmentation. A Wav2Vec 2.0 acoustic model was trained with the Common Phone to perform phonetic symbol recognition and validate the quality of the generated phonetic annotation. The architecture achieved a PER of 18.1 % on the entire test set, computed with all 101 unique phonetic symbols, showing slight differences between the individual languages. We conclude that Common Phone provides sufficient variability and reliable phonetic annotation to help bridging the gap between research and application of acoustic models.