Deepgram vs AssemblyAI vs Whisper
Speech recognition’s core matchup: the two API leaders against the open-weights reference.
Verdict
Nova-3 wins on accuracy+speed, AssemblyAI on accent robustness, Whisper on self-hosting economics.
Side-by-Side Comparison
| # | Model | WER (lower = better) | Speed (RTF) | First Partial | Cost / 1K Min | Mode |
|---|---|---|---|---|---|---|
| 1 | Nova-3 (Speech-to-Text) Deepgram | 4.8% | 110x | 0.18s | $4.30 | Streaming & Batch |
| 3 | Universal-3.5-Pro AssemblyAI | 5.1% | 95x | 0.22s | $6.50 | Streaming & Batch |
| 7 | Whisper large-v3 turbo OpenAI / Open Source | 5.4% | 85x | 0.35s | $6.00 | Batch Async |
Model-by-Model
Nova-3 (Speech-to-Text) — Deepgram
Deepgram Nova-3 leads accuracy at 4.8% WER with 110x real-time speed and 0.18s first-partial latency — $4.30/1K min.
Universal-3.5-Pro — AssemblyAI
AssemblyAI Universal-3.5-Pro matches at 5.1% WER with streaming+batch flexibility at $6.50/1K min — strongest on English-accent robustness.
Whisper large-v3 turbo — OpenAI / Open Source
Whisper large-v3 turbo (5.4% WER) is the open-weights baseline — slower (0.35s TTFP, batch only) but self-hostable and effectively free at scale.
FAQ
What is the most accurate speech-to-text API?
Deepgram Nova-3 at 4.8% WER on the AA-WER v2 composite, followed by ElevenLabs Scribe v2 (4.9%) and AssemblyAI Universal-3.5-Pro (5.1%).
Is Whisper still worth using?
For self-hosted pipelines and privacy-critical transcription, absolutely — open weights with 5.4% WER. For hosted APIs, Deepgram and AssemblyAI are faster with better accuracy.
Which STT is cheapest?
Distil-Whisper at $1.20/1K min and Parakeet TDT v3 at $1.80 are cheapest overall. Of this trio, Deepgram Nova-3 ($4.30) undercuts AssemblyAI ($6.50) and Whisper API ($6.00).