Which speech-to-text API is the most accurate?

On the Hugging Face Open ASR Leaderboard, the most accurate cloud API is OtterAI/speech-v3.1, with an average word error rate of 3.88 percent ¹. In second place is zoom/scribe_v2_pro at 4.08 percent ². Lower is better: it is the share of words the API gets wrong. The reading is the leaderboard's version of October 17, 2026, taken on October 20, 2026, and it ranks 13 cloud APIs next to open-weight models.

Otter.ai leads the cloud APIs, and Zoom is right behind

The average covers nine public English test sets: read audiobooks, meetings, earnings calls, podcasts and videos, European Parliament speeches, noisy speech and Indian English conversations. Every model is scored on the same audio, so the averages are directly comparable.

Microsoft's Azure AI Speech comes third at 4.49 percent ³. Modulate Velma Transcribe follows at 4.53 percent ⁴, then ElevenLabs Scribe at 4.55 percent ⁵.

Showing 8 of 13 rows. See every row in the dataset.

The top seven APIs are tightly bunched

The gap between first and second place is small, and the next several APIs stay close behind. The seventh cloud API, reson8/resonant-1, averages 4.64 percent ⁷, less than one point above the leader. In between sits sophea/asr-k1 (preview) at 4.61 percent ⁶, along with the APIs named above.

Further down, the spread widens. OpenAI's gpt-transcribe averages 5.49 percent ⁹, and the last cloud row, meta/muse-voice-transcribe, averages 5.75 percent ¹⁰.

Leading on average does not mean leading on every test

Otter has the lowest average, but it does not win every column. Zoom has the lower error rate on five of the nine test sets: meetings (AMI), Earnings-22 calls, SPGISpeech, noisy speech and Indian English conversations. Otter has the lower rate on the other four: both LibriSpeech sets, GigaSpeech and VoxPopuli. Modulate has the lowest LibriSpeech clean score of any cloud API.

The full table shows these columns for every row. AssemblyAI's universal-3-5-pro, at 4.87 percent on average ⁸, shows why: it sits behind the leaders overall, yet its AMI meeting score is the weakest of the top eight cloud APIs. Averages smooth over such gaps.

If your audio looks like one test set (a meeting, say, or a finance call), read that column instead of the average.

What the numbers do and do not measure

• Public sets only. The leaderboard's own Average WER also includes two averages over private recordings that are not published, so its ranking can differ slightly from this public-set average.
• Not dictation. The audio is audiobooks, meetings, earnings calls, podcasts, parliament speeches, degraded speech and conversations, not a person dictating an email.
• Normalized text. Capitals and punctuation are removed before scoring, so the score says nothing about formatting quality.
• No speed or price. Cloud APIs have no speed figure here, and the table has no price, latency or privacy information.
• Preview rows. One row is labeled a preview, so its score may change. The values are read from the leaderboard on a schedule and will update.

What this means for choosing an API

For batch transcription of clean speech, the leaders are close enough that price, language support, data handling and speed should decide. Test two or three candidates on your own recordings, because a difference between leaders on public audio can reverse on yours.

If you need an API for a product, compare the cloud rows with the open models too: the best open-weight model beats five of the cloud APIs on average, as our page on open-source models against paid APIs shows. For noisy audio, see which model handles noise best.

Related pages

• The full dataset, with every row and its source
• Market overview of speech-to-text APIs and all products compared
• Speech recognition accuracy across studies
• How speech recognition benchmarks work and the speech recognition overview