NewsShorts-Bench results: open AI models on real news broadcasts (October-November 2026)
Authors/Creators
Description
Results tables for 12 stages of NewsShorts-Bench, a series of reproducible benchmarks that test open AI models on the steps of turning real news broadcasts into short clips: speech recognition, word alignment, speaker diarization, on-screen name-bar OCR in 13 languages, active speaker detection, story segmentation, segment roles, highlight selection, misleading-clip detection and repair, translation faithfulness, and YouTube most-replayed proxies.
Each stage asks one question, runs several open systems on the same material, and reports one headline metric. index.csv lists every stage with its question, verdict, metric, method, hardware, run date and the URL of the full write-up; results/<stage>.csv has one row per system.
The numbers are exported from the published write-ups at everylocalai.com/benchmark, which include the full method, per-video notes and caveats. Source videos are not included.