Ursa

Audio representations, phonemes and words

Ursa first learned audio representations from more than a million hours of unlabelled speech across 50 languages. A supervised acoustic model then mapped those representations to phoneme probabilities. A language model converted the probabilities into likely words.Speechmatics: Introducing Ursa, 7 March 2023

Its domain study compared performance across medical, financial, technical, political and construction recordings. The study addressed generalization without custom domain adaptation, rather than presenting those domains as five separately released models.Speechmatics: Ursa and domain generalization

Practical guidance. The two-billion-parameter figure applies to the self-supervised component. Treating it as an end-to-end system size would omit the acoustic and language components.Speechmatics: Introducing Ursa, 7 March 2023