Device and accent diversity for speech models that generalise

A speech model trained on clean audio from one device and one accent learns that device and that accent — and falls over on everything else. Generalisation is a data-design decision, made before the first recording.

Accent diversity in recruitment

We recruit across regional accents and socio-linguistic backgrounds as part of the sampling matrix, so the corpus reflects how a language is actually spoken across regions — not just its prestige variety.

Device diversity on purpose

Capture spans Android and iOS smartphones, laptop and desktop microphones, headset mics and field recorders. Each colours the audio differently; covering the range is what lets a model handle real-world input.

Tag it so you can balance it

Device type and environment category are recorded as metadata per record. That's what makes the diversity usable — you can balance, split or stress-test against it later.

It starts with collection

This is a collection decision first and a modelling decision second. Build the diversity in, and generalisation follows.