AssemblyAI Universal-2

Universal-2 is AssemblyAI's workhorse speech-to-text model, expanded to 99 languages in August 2025. It's a 600M-parameter Conformer RNN-T model trained on 12.5 million hours of multilingual audio, with a second-stage neural formatting pass that handles punctuation, casing, and text normalization. Its differentiator is bundled speech intelligence: sentiment analysis, topic detection, entity recognition, and PII redaction run in the same API call as transcription, rather than as separate services.

The honest picture: AssemblyAI is API-only with no open weights, so there's no self-hosting option. Independent latency benchmarks have measured it slower than Deepgram on streaming (761ms P50 in Coval's 2026 test). The newer Universal-3 Pro beats it on raw accuracy for its 6 core languages, so Universal-2 is now the value play ($0.15/hr) rather than the accuracy leader — pick it for broad language coverage and the bundled intelligence features, not for winning WER shootouts.