Australian English is a specific consideration

General-purpose speech-to-text models are trained predominantly on American English audio. Australian English, with its distinctive vowel sounds, vocabulary, and speech patterns, performs less well on these models than on models specifically trained or fine-tuned on Australian data.

For Government deployment in Australia, Agencies should ask vendors specifically about Australian English performance data, not just general accuracy claims. The gap between a model trained on Australian data and a generic English model can be significant, particularly for regional accents and APS-specific terminology.

A vendor that cannot produce Australian English benchmarks is selling a model that was not designed for the Agency’s actual operating environment.