Built on how people really speak

The Saffar Kernel Atlas

We analysed 111 million words of real conversation across 13 languages and ranked every word by how often it is actually heard, after grouping word forms into their base word.

The result is a kernel for each language: the smallest set of words that makes up 80%, 90% and 95% of everyday speech.

Tested on speech it had never seen

A kernel is only useful if it holds up in new conversations. We built each kernel on part of the data and tested it on the rest: the 90% kernels covered between 89.8% and 90.0% of the unseen speech.

A memory model for planning

To turn kernel sizes into a daily plan, Tooti uses a model of how quickly words are learned and forgotten. It estimates how many minutes a day each goal needs, and your own data refines it every week.

Honest claims only

We never promise fluency or a guaranteed result. Every number on this site comes from the atlas or the planning model and is stored with its source.

We will publish learner outcomes only after a pilot study with a pre-registered analysis.