Universal-2
AssemblyAI's multilingual speech model — transcription plus audio intelligence in one API
About
Universal-2 is AssemblyAI's flagship speech-to-text model, released October 2024. It uses a 600M-parameter RNN-T architecture: the encoder was pre-trained on 12.5 million hours of unsupervised multilingual audio, then the full model fine-tuned on 300,000 supervised hours. Beyond transcription it bundles the audio-intelligence layer AssemblyAI is known for — speaker diarization, summarization, sentiment analysis, entity detection, topic detection, PII redaction, and a disfluency toggle for filler words. Async file API plus real-time WebSocket streaming. Pricing is modular: $0.15/hr base transcription with intelligence features as add-ons, which can roughly triple the bill when everything is on. It is a strong default for products that do something with audio beyond just transcribing it.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: accuracy on par with best-in-class
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Released Oct 2024; 600M-param RNN-T, 12.5M-hour unsupervised pretraining per ASR literature; 99-language coverage and bundled audio intelligence corroborated in SDK docs; Modular pricing ($0.15/hr base + add-ons) captured honestly.