Medical STT benchmark v4 reshuffles rankings

Standard Word Error Rate (WER) is a dangerous metric for medical AI when it weights filler words equally with life-critical drug names.

• –Medical-WER (M-WER) reveals that top general models often butcher drug names, with error rates 2-5x higher than other categories.
• –Microsoft's VibeVoice-ASR 9B (#3) beating its closed MAI-Transcribe-1 (#11) by 1.7 points highlights the power of LLM-backed transcription.
• –Qwen3-ASR 1.7B emerges as the best small open-source model, delivering near-Gemini performance at 14x the speed of larger models.
• –Deepgram Nova-3 Medical holds its ground as the fastest cloud API, completing files in just 13 seconds without compromising accuracy.