Ambience described a change to its clinical speech recognition pipeline from a mini batch ensemble to a rolling window ensemble, which keeps recent audio in scope so a reconciliation model can revise earlier transcript segments as more audio arrives. Ambience reports keyword error fell 16 percent against the previous design on a 226 encounter evaluation spanning 28 specialties and 15 clinicians, at the cost of a median processing cycle roughly 40 percent longer.
Our readErrors on drug names, dosages and negations are what make an ambient note unsafe to sign, which is why keyword error is the figure to watch. The evaluation is the vendor's own and spreads 226 encounters across 28 specialties, so ask whether the gain holds in your highest volume specialties.