Curiosity · AI Model
AssemblyAI Universal-2
Universal-2 is AssemblyAI's flagship ASR model. It is tuned for batch transcription quality — achieving state-of-the-art WER on English meetings and media — and integrates with LeMUR, AssemblyAI's LLM layer, for one-call summaries, chapters, sentiment, and Q&A on top of the transcript.
Model specs
- Vendor
- AssemblyAI
- Family
- Universal
- Released
- 2024-10
- Context window
- 1 tokens
- Modalities
- text, audio
- Input price
- n/a
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- State-of-the-art English WER on meeting-style audio
- LeMUR LLM features on top of the transcript — summaries, chapters, sentiment, Q&A
- Speaker diarisation, auto-chapters, PII redaction built in
- Simple REST API with usage-based pricing
Limitations
- Batch-first — streaming latency lags Deepgram Nova-3
- Multilingual coverage smaller than Whisper large-v3
- Closed API — no self-hostable weights
- LeMUR features depend on AssemblyAI's own LLM pipeline, not your choice of model
Use cases
- Podcast and media transcription with chapters
- Meeting notes with speaker diarisation and summaries
- Content moderation and topic detection
- Sales-call analytics and Q&A over recordings
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| English WER (meeting audio) | ≈6.6% | 2024 |
| Streaming p95 latency | <500 ms | 2024 |
Frequently asked questions
What is AssemblyAI Universal-2?
Universal-2 is AssemblyAI's flagship speech-to-text model, released in October 2024. It targets batch transcription quality (state-of-the-art English WER) and integrates with LeMUR, AssemblyAI's LLM layer, for downstream analysis.
What is LeMUR?
LeMUR is AssemblyAI's LLM layer that operates on transcripts — it produces summaries, chapters, sentiment, action items, and supports Q&A over the recording in a single API call on top of Universal-2 output.
How does Universal-2 compare with Deepgram Nova-3?
Universal-2 leads on batch transcription quality and bundles LLM analytics; Nova-3 leads on streaming latency and voice-agent latency budgets. Pick Universal-2 for podcast and meeting workflows; Nova-3 for real-time phone agents.
How is AssemblyAI priced?
AssemblyAI bills per second of audio for transcription, plus per-token fees for LeMUR LLM features. Volume discounts and enterprise plans are available.
Sources
- AssemblyAI — Universal-2 — accessed 2026-04-20
- AssemblyAI — Docs — accessed 2026-04-20