Fastest Speech-to-Text 2026: Latency Benchmarks
"Fastest STT" is three different races, and vendors quote whichever one they win. Before the numbers, know which race you are running:
- Time to first token — when the first words appear. Decides whether the agent feels alive.
- Time to final segment — when words stop changing. Decides when you can act on the transcript.
- Endpointing — how fast the system detects you're done talking. Often bigger than both above (see our Flux vs VAD breakdown).
Streaming Latency Leaderboard (Independent, September 2026)
Daily third-party testing (Coval, fixed public audio, 30-day averages — lower is better):
| Model | Time to first token | Rank |
|---|---|---|
| AssemblyAI Universal-3.5 Pro | 1,034ms | #1 |
| Deepgram Flux | 1,083ms | #2 |
| Deepgram Flux Multilingual | 1,170ms | #3 |
| NVIDIA Parakeet TDT 0.6B v3 | 1,213ms | #4 |
| Whisper Large v3 | 1,293ms | #5 |
| Deepgram Nova-3 | 1,426ms | #6 |
Surprises: a free open model (Parakeet) beats Nova-3 to first token, and Whisper Large v3 — written off as slow — sits mid-pack on this metric.
Final-Segment Speed: A Different Winner
| Model | Time to final segment | Rank |
|---|---|---|
| ElevenLabs Scribe v2 Realtime | 123ms | #9 of 24 |
Scribe v2 Realtime is the finals-speed story — but remember its 2,162ms first token (#20 of 22). Slow to start, fast to finish. Whether that suits your agent depends on whether users pause mid-thought.
Vendor Claims, Translated
| Claim | What it means |
|---|---|
| Flux "<200ms endpointing" | Turn-detection buffer, not transcription — real, and the right metric for conversations |
| Nova-3 "<300ms streaming" | Transcription latency on clean audio; first tokens in tests land ~1.4s |
| Scribe v2 Realtime "sub-150ms" | Finals, confirmed independently (123ms); first tokens are the catch |
| Universal-3.5 Pro "~300ms turn detection" | Agent-API figure; first token leads the independent board |
| Smallest Pulse "<100ms" | Time-to-first-byte claim — verify against the TTFT board above before believing |
The Honest Verdict
- Snappiest-feeling agent today: Universal-3.5 Pro Realtime or Flux — they top first-token boards and handle endpointing.
- Fastest finals: Scribe v2 Realtime, if your UX tolerates slow starts.
- Fastest free option: Parakeet TDT v3 — first tokens faster than Nova-3 at $0 and 600M params. See Parakeet vs Whisper.
- Biggest latency lever overall: endpointing, not STT. A 200ms transcriber behind a 1,000ms VAD is a 1,200ms agent.
FAQ
What is the fastest speech-to-text API? By independent first-token testing (Sep 2026): Universal-3.5 Pro (1,034ms), then Flux (1,083ms). By vendor claims: Pulse (<100ms TTFB, unverified independently).
Does STT latency even matter most? Usually no — endpointing and LLM round-trips dominate. Fix turn detection before chasing STT milliseconds.
Related comparisons & reviews
Keep reading
Voice Model Deep Dives
Best Speech-to-Text API 2026: Benchmarks & Prices
All 8 leading STT APIs compared: Nova-3, Universal-3.5 Pro, Scribe v2, Gemini 3.5, MAI, Soniox, GPT-4o. Prices, WER and picks.
Voice Model Deep Dives
ElevenLabs Scribe v2 Review 2026: Realtime & Pricing
Scribe v2 at $0.22/hr plus a sub-150ms Realtime model. Benchmarks, pricing vs Deepgram and AssemblyAI, and the diarization catch.
Voice Model Deep Dives
Gemini 3.5 Transcribe Review: Smart STT Tested (2026)
Google’s Aug 2026 model polishes ramblings into formatted text. 85+ languages, smart transcription, and how it compares to Nova-3.