Speech models struggle most when speakers come from diverse linguistic backgrounds — and nowhere is that more costly than navigation and emergency dispatch. New from the Together Research Frontier Agents team: SF Streets, a benchmark stress-testing named entity recognition across 15 state-of-the-art models. → 39% average error rate on street names → Non-English speakers: 18% lower accuracy → Mis-transcriptions land you 2.4 miles off target The fix: cross-lingual style transfer. Fewer than 1,000 synthetic samples → 60% relative improvement on Whisper-Large. SF Streets and US Streets datasets releasing publicly. Read more and find the paper (links below)

Blog: arXiv: