Fact · Voice and AI
32.5percent
On the 2012 Switchboard task, the move from a GMM-HMM baseline at 27.4 percent word error rate to a deep network at 18.5 percent is a relative reduction of about 32.5 percent.
Checked against its source on October 17, 2026
- What it measures
- A relative reduction divides the drop in error rate by the starting error rate. The absolute drop is 8.9 percentage points.
- How it is calculated
- (27.4 - 18.5) / 27.4 = 0.325, about 32.5 percent relative reduction
- What it does not show
- One benchmark from 2012. Modern recognizers are far more accurate, so the size of this gain should not be read as a current figure.