Why everyone else is making speech smoother

A note on what we're building, and why it runs against the grain.

The large speech models that power dictation and voice assistants are trained to do one thing very well: turn messy human audio into clean text. They smooth out the stumbles, drop the false starts, and hand back something tidy. For most products that is exactly right.

For a child in speech therapy, it is exactly wrong. The block, the repeated sound, the drawn-out syllable, those are not noise to be cleaned away. They are the signal. They are what the therapist needs to see.

So we build the opposite

Qull trains its own model on disfluent speech from real practice. Instead of erasing the stumble, it holds onto it and measures it, so progress becomes something a family and a clinician can actually watch over time.

More to come as we publish how this works.