Universal-2

Universal-2 is AssemblyAI's previous-generation speech-to-text model, announced at the end of October 2024 and still sold alongside the 2026 flagship rather than retired behind it. It is the unusual case of an older model that is broader than its successor: it transcribes more than 99 languages, where Universal-3.5 Pro covers 18, and it costs 0.15 dollars per hour of audio against the flagship's 0.21. The model's stated design goal was not raw word accuracy but the parts of a transcript a reader actually notices when they are wrong: names, numbers, and formatting. On the manufacturer's own comparison it records a 13.87% Jaro-Winkler error rate on proper nouns, 10.06% word error rate on text formatting and 4.00% on alphanumerics — figures the company puts ahead of the contemporary versions of Deepgram Nova-2, OpenAI Whisper Large-v3, Microsoft Azure Batch v3.1, Amazon and Google Latest-long on all three counts except alphanumerics, where Whisper is narrowly better. These are the vendor's measurements of its rivals and should be read as such. Universal-2 also supports code switching and keyterms prompting, in which the caller supplies up to 200 domain words — product names, drug names, personal names — that the model is told to expect. AssemblyAI now positions it explicitly as the fallback for languages the flagship does not cover and as the choice for high-volume, price-sensitive batch work. The trade the buyer is being asked to make is stated plainly by the manufacturer: newer and sharper on eighteen languages, or older, cheaper and workable on ninety-nine.

▸Videos

No videos yet.

▸Reviews