Three new transcription models on medical audio
Azure MAI-Transcribe-1.5, OpenAI gpt-transcribe and Google Gemini 3.5 Transcribe joined the same sealed 1,513-clip board.
30 systemsDrug & dosage scores
See what changed →Updates & field notes
What changed—and what we measured.
Dated medical speech model tests, benchmark changes and product evidence. We publish when there is a useful result, not to fill a content calendar.
Latest
Medical speech, tested as it changes.
Each update links the result back to the complete board, method and product decision it can—and cannot—support.
The current medical speech benchmark
Overall WER, medical-term error, drug-name error and dosage F1 on the same clinical audio.
Version 71,513 clips
Open the current board →Models, methods and research
Open medical speech models, benchmark methodology and note-generation research in one archive.
Browse the research →Publishing standard
Evidence before announcement.
Every Omi update should make the finding, denominator, method and limitation easy to inspect.
FindingLead with what changed in the world, not a launch adjective.
MethodShow the same input, scorer and version behind the comparison.
CaveatState what the result does not measure or generalize to.
ActionGive the reader a board, model, test or integration to inspect next.
Test the result on your audio.
The benchmark narrows a shortlist. Your workflow makes the decision.