Best English Speech-to-Text API (2026): independent benchmark
English speech-to-text accuracy, measured. 50 English read-speech clips from FLEURS, every provider routed through the same gateway and scored identically on 2026-06-03. Lower WER is better.
OpenAI GPT-4o Transcribe posted the lowest WER at 2.4%, with Alibaba Qwen3-ASR (2.6%) close behind. With n=50 the confidence intervals are about plus or minus 2 points, so the top four are a statistical tie. The spread below the top is real: the board runs from 2.4% to 13.2%.
English STT leaderboard
50 English clips (FLEURS), measured 2026-06-03 through the Speko gateway, loudness-normalized to -16 LUFS. WER measured on English; latency and list price from the same gateway setup on the English board. Lower is better.
| Provider / model | WER | p50 latency | List price |
|---|---|---|---|
| OpenAI GPT-4o Transcribe | 2.4% | 1,084 ms | $0.006/min |
| Alibaba Qwen3-ASR | 2.6% | 2,195 ms | - |
| ElevenLabs Scribe v2 | 2.9% | 1,353 ms | $0.0067/min |
| xAI Grok STT * | 4.8% | 996 ms | - |
| Cartesia Ink-2 | 6.1% | 966 ms | $0.0022/min |
| Gradium | 13.2% | 2,528 ms | - |
* Returns an empty transcript on low-level audio - it dropped ~half of un-normalized quiet clips. 4.8% is on loudness-normalized input.
How we measured
- Dataset: 50 English read-speech clips from FLEURS.
- Scoring: mean word error rate (WER), lower is better. Audio is loudness-normalized to -16 LUFS before scoring so input-gain handling does not contaminate the accuracy column.
- Every provider is measured the same way: through the Speko gateway (POST /v1/transcribe, provider pinned), from a single location.
- Latency is p50 end-to-end and includes the gateway hop for every provider. Cost is the provider list price per minute of audio; a dash means no public list price.
- Run date: 2026-06-03. n=50, so confidence intervals are about plus or minus 2 points: the top four are a statistical tie.
Full interactive table, every territory, and the complete methodology: benchmarks.speko.ai
Use the winner without lock-in
The best English provider today is one benchmark run away from being second best. Speko is one API in front of every provider on this table: it routes each request to the measured-best provider for your language and fails over automatically when a provider degrades. No per-vendor integration, no migration when the leaderboard flips.
curl -X POST https://api.speko.dev/v1/transcribe \
-H "Authorization: Bearer $SPEKO_API_KEY" \
-H "Content-Type: audio/wav" \
-H "x-speko-intent: {\"language\":\"en\"}" \
--data-binary @call.wavimport { Speko } from '@spekoai/sdk';
import { readFile } from 'node:fs/promises';
const speko = new Speko({ apiKey: process.env.SPEKO_API_KEY! });
const audio = await readFile('./call.wav');
const { text, provider, confidence } = await speko.transcribe(audio, {
language: 'en',
});FAQ
What is the most accurate English speech-to-text API?
On Speko's 2026-06-03 FLEURS benchmark (50 English clips), OpenAI GPT-4o Transcribe posted the lowest WER at 2.4%, followed by Alibaba Qwen3-ASR at 2.6%. With n=50 the confidence intervals are about plus or minus 2 points, so the top four are a statistical tie.
Does ElevenLabs support English speech-to-text?
Yes. ElevenLabs Scribe v2 scored 2.9% WER on our English run.
Which English STT API is fastest?
Cartesia Ink-2 had the lowest p50 latency on our run at 966 ms. Every latency number includes the gateway hop, so the comparison is like-for-like across providers.
Which English STT API is cheapest?
By list price, Cartesia Ink-2 at about $0.0022/min. OpenAI GPT-4o Transcribe lists at $0.006/min and ElevenLabs Scribe v2 lists at $0.0067/min. Providers without a public per-minute list price show a dash.
Why does the xAI row carry a note?
Returns an empty transcript on low-level audio - it dropped ~half of un-normalized quiet clips. 4.8% is on loudness-normalized input.