At 10:30 last night I ran our platform’s new ASR update on a 12-minute ortho consult; edit time dropped from 8:42 to 8:19 and WER from 6.1% to 5.6% — tiny, but it adds up over a shift. What are your per-10-minute benchmarks since the patch, and are you seeing similar deltas on acronyms and drug names?
I’m at 6:18 per 10; WER 5.9→5.4 since the patch. @emma-gar19 The one thing that moved meds for us was uploading our site formulary as phrase hints (CSV) to the custom vocab; acronyms are better but still slip on fast dictation — anyone tweak biasing for all-caps?
Quick example: I got a bigger lift by turning on “numeric normalization: strict” so [redacted] stays numeric and by forcing sub-5-char acronyms to stay uppercase (ACL, OA), which cut med and ortho slips for us. Tiny caveat: it can over-preserve things like “q.d.” if your punctuation cleaner is aggressive. @emma-gar19 have you tried that setting?
On your 12‑minute consult, nice 8:42→8:19; what moved meds for us was turning on n‑best=5 and weekly mining the confusions, then modest boosts on the top 50 drug names plus brand/generic pairs. Small caveat: over‑boost will hallucinate in ortho, so cap weights and feed the chart header/problem list as context: https://cloud.google.com/speech-to-text/docs/contextual-biasing. What backend are you on?
Seeing a small bump here too: turning on word‑confidence highlighting and adding a tiny clinic lexicon (surgeon names + local implant brands) shaved about 20s per 10 min and killed my ‘AC joint’ vs ‘A.C.’ slips. If you can, let the model auto‑boost from the day’s schedule — pit‑stop seconds, but they add up; @beaver, have you tried schedule boosts for acronym‑heavy clinics?
I saw a tiny hinge, big door effect by switching numeric normalization to digits+units and locking all‑caps for known acronyms; fewer dose fixes and fewer ACL/MCL flips. Keeping hyphens in measurement tokens (e.g., 5.5‑mm) cut implant-size edits and saved about 15s per 10 min here, though it only sticks if your export doesn’t recase/punctuate on save. Do you have a toggle for per‑note “context seeding” from the patient med list, or is that off‑limits in your setup?