What is the most accurate speech-to-text model?

No single model is the most accurate speech-to-text model for everyone. On public English tests, the best systems form a near tie, and the leader changes with each new release. In the most recent leaderboard snapshot we could check, the top average word error rate was about 5.3 to 5.4 percent ^{1 2}, on English tests that mix read and spontaneous speech. That number says little about your voice, your microphone or your room. The accurate answer for you is the model with the lowest word error rate on recordings like yours.

The top models on public English tests sit within a fraction of a point

The Open ASR Leaderboard is the most widely used public comparison. In the English short-form table of the Open ASR Leaderboard paper, Cohere Labs Transcribe is marked as open and had an average word error rate of 5.42 percent ². Zoom Scribe v1, which the table marks as not open, scored 5.47 percent ³. A later mirror of the results, dated May 6, 2026, lists IBM Granite Speech 4.1 2B first among open models, at 5.33 percent ¹.

A gap of less than a tenth of a point is small enough that a different test set could reorder these systems. Treat the top group as one tier. The order is a snapshot, not a ranking that will hold.

Our guide to speech recognition benchmarks explains what the scores measure and what they leave out, and the speech recognition overview covers the subject as a whole.

Clean read speech is the easiest case for every model

The leaderboard's English tests mix read audiobooks and prepared talks with spontaneous speech such as meetings and earnings calls. Every system does better there than it will in daily use. Harder audio raises the error rate sharply. A 2026 study of commercial systems found that the leading services reached about 1 to 2 percent word error rate on mild dysarthria, a speech condition that affects clarity ⁴. On severe dysarthria, every commercial system exceeded 51 percent word error rate ⁵.

Atypical speech is an extreme case, but it shows that accuracy belongs to a model and a speaker together. Our guide to speech recognition and accents covers the accent side of the same problem.

The best model depends on the language

Public English scores say little about other languages. In the Whisper paper's FLEURS results for the large-v2 model, the word error rate was 3.0 percent in Spanish ⁶ and 99.0 percent in Telugu ⁷. The same model can be excellent in one language and close to unusable in another, so a language-specific test matters more than any English leaderboard. For the language picture, see our guide to Whisper supported languages and accuracy.

Speed and accuracy are separate choices

The most accurate model is not always the one you want. A large model may be more accurate but slower, and a smaller model may be fast enough to feel instant. The leaderboard reports speed as a separate measure for that reason ⁸. For how delay affects dictation, see our guide to real-time speech recognition.

Lower word error rate does not always produce a better reading experience either. In a 2017 study of automatic captions for deaf and hard-of-hearing viewers, word error rate correlated only 0.11 with user ratings ⁹. That is one study, but it is a reminder that the most accurate score is not always the most useful text.

It depends on your audio, so test it

The answer for most people is a short list, then a test. Pick the two or three systems at the top of the track that matches your language. Record ten to twenty minutes of the speech you will actually use, with the microphone and room you will use. Transcribe it by hand as the reference, decide in advance whether fillers and punctuation count, and score each system with the same method. Word error rate is the metric to use. Our question on what a good word error rate is explains how to judge the number.

For dictation on your own computer, the free built-in tools are often enough. You can read about the options in our guide to voice to text on Mac, and you can try a clip in the browser with our free audio to text tool. Yap, which we make, is an on-device dictation app for Mac and Windows, so you can test it with the same recordings. The download page has the installers.

What to do

1. Choose a track that matches your language and audio type, not the overall average.
2. Shortlist the top two or three systems, and drop any that are far behind.
3. Test each on your own recordings with one scoring method.
4. Compare word error rate and delay, and choose the system that is best on both for your use.
5. Recheck the choice every few months, because the leaders change.

For why published numbers disagree, see why speech-to-text accuracy claims disagree.