NoiseScribe demo console: Interview — Dr. Amelia Reyes

Most of the failure cases we saw weren't accent-related at all. They were rooms — HVAC, a projector fan, two people talking over each other.

And the gating happens before the model, not after?

Before. We run spectral subtraction and an RNN suppressor on the buffer, then hand a clean 16-kilohertz stream to the transcriber. Nothing downstream has to guess.