Which speech-to-text API
is the most accurate?
Record your voice or upload an audio file. We'll transcribe it with two providers. You compare, pick the best and learn who's behind it.
Live speech-to-text API leaderboard
| Rank | Provider | ELO score | Votes | 95% CI |
|---|---|---|---|---|
| 1 | Gladia (Solaria-3) | 1576 | 340 | ±43 |
| 2 | ElevenLabs (Scribe v2) | 1560 | 340 | ±43 |
| 3 | AssemblyAI (Universal-3 Pro) | 1549 | 340 | ±43 |
| 4 | Mistral (Voxtral Mini) | 1473 | 340 | ±43 |
| 5 | Speechmatics (Enhanced) | 1468 | 340 | ±43 |
| 6 | Deepgram (Nova 3) | 1373 | 340 | ±43 |
Based on 1,020 blind comparisons. Last updated 2026-10-29. See the full leaderboard.
How the speech-to-text API comparison works
Compare STT is a blind, head-to-head test of six commercial speech-to-text APIs: Gladia (Solaria-3), ElevenLabs (Scribe v2), Deepgram (Nova 3), AssemblyAI (Universal-3 Pro), Speechmatics (Enhanced) and Mistral (Voxtral Mini).
You record or upload up to two minutes of audio. Two providers transcribe the same audio file in parallel, each through its official SDK or REST API, with no preprocessing, normalization or prompt engineering. You see the results as Model A and Model B and vote for the more accurate transcript, or call a tie. Provider names are revealed only after you vote.
Pairs are not picked purely at random. The system chooses among the least-played of the 15 possible provider pairs, so every matchup builds up a similar number of comparisons over time. A coin flip decides which provider appears as Model A.
Votes feed an ELO rating, the pairwise system used to rank chess players. Every provider starts at 1,500, the K-factor is 32, and ratings are recomputed in one chronological pass over every vote. The leaderboard shows each provider's ELO score, vote count and 95% confidence interval. Rankings stay blurred until the results reach statistical significance.
Each vote stores only the two providers compared and the winner. Audio is deleted as soon as transcription finishes, and no transcripts or user accounts are kept. Single-use signed match tokens, a per-session vote cap, a minimum delay before voting and rate limits protect the results from automated voting.
The test has known limits. The audio is submitted by users, so the dataset is not controlled, and each voter decides what "better" means. ELO reflects aggregate human preference on real-world audio, not word error rate on a fixed dataset. For word error rate results on public datasets, see Gladia's speech-to-text benchmarks.
Gladia sponsors the project and pays for every API call. No provider supplied free credits, and the full source code is public on GitHub, so anyone can audit how votes are counted.
Speech-to-text API comparison FAQs
On Compare STT, accuracy is measured through blind community votes, not a single lab benchmark. Rankings change as users compare transcriptions side by side. Check the live leaderboard for current ELO scores across Gladia, ElevenLabs, Deepgram, AssemblyAI, Speechmatics, and Mistral.
You upload or record audio. Two speech-to-text providers transcribe it anonymously and you vote for the more accurate one. Votes feed an ELO ranking system that updates the public leaderboard.
ELO is a pairwise rating system originally used in chess. Each blind vote shifts provider ratings up or down based on expected vs. actual outcomes. See the full methodology and current standings on the leaderboard.
Six providers are in the arena: Gladia (Solaria-3), ElevenLabs (Scribe v2), Deepgram (Nova 3), AssemblyAI (Universal-3 Pro), Speechmatics (Enhanced), and Mistral (Voxtral Mini). Read about each provider's setup in the methodology.
Test each API on audio that matches your real use case: the same languages, accents, audio quality and length you will run in production. Compare word error rate on a labelled sample, then latency, language coverage, features such as diarization, and price per hour.
Compare STT adds a fast first pass. Upload or record up to two minutes of your own audio, get two anonymous transcripts from different providers, and vote for the better one before the names are revealed.
Word error rate counts the share of words a model gets wrong against a human reference transcript on a fixed dataset. An ELO ranking comes from pairwise human votes: each time one transcript is preferred over another, the winning provider gains rating points and the loser drops.
Compare STT uses ELO, so its leaderboard reflects which transcripts people prefer on real-world audio, including punctuation and formatting. For word error rate on public datasets, see Gladia's speech-to-text benchmarks.
A sponsored test is only as credible as its method. The checks that matter are whether voting is blind, whether every provider gets the same audio, and whether the code is public.
On Compare STT, provider names stay hidden until after the vote, every provider receives the same audio file through its official API with no preprocessing, and pairs are balanced so no matchup is over-represented. Gladia pays for every API call, no provider supplied free credits, and the full source code is on GitHub for anyone to audit.
Yes. Provider names and logos are hidden until after you vote. You only see two anonymous transcripts labeled Model A and Model B, reducing brand bias. Learn more in the methodology or read about the project.
Rankings update continuously as new votes are submitted. ELO scores are recomputed from the full vote history, so every comparison can shift provider standings. See the live leaderboard for current results.
Compare STT is sponsored by Gladia, a speech-to-text API company. The benchmark is open source and community-driven; see the methodology for full transparency on scoring and provider settings.
See the live leaderboard or read the full methodology.