Which speech-to-text API is the best?

There is no single best speech-to-text API, and anyone who names one without hearing your audio is guessing. Independent benchmarks show that the strongest systems are close on clean English. The best API for you is the one that makes the fewest errors on your recordings, arrives fast enough for your use, and fits your budget and data terms. The speech-to-text API guide has the full comparison, and the overview of speech recognition explains the technology underneath.

On clean English, the leading systems are close

The Open ASR Leaderboard tests open-source and proprietary systems on shared test sets, and reports the average word error rate across them. The table shows three systems from its English short-form track.

The top rows are separated by a small margin, and the gap to Whisper Large v3 is modest too. That is the sense in which the leaders are close. The score is also measured on normalized text, with punctuation, casing and fillers removed ⁴. The table therefore says nothing about punctuation, capital letters, or whether a dictation tool keeps the word "um". Those are the parts of a transcript that people edit by hand, so test them yourself.

Live captions need a different test

Ranking changes once speech arrives in real time. In Artificial Analysis's 2026 streaming benchmark, the first partial transcript from Deepgram Flux had a word error rate of 7.4 percent, while ElevenLabs Scribe v2 Realtime's first partial had 3.6 percent ⁵. A batch leaderboard does not measure that. Streaming also tends to be harder for recognizers: a 2024 study found significantly lower quality for streaming speech recognition ⁶. If you are building live captions or a voice agent, benchmark the streaming endpoint. The guide to real-time speech recognition explains the mechanics.

Hard speech and invented text can change the answer

A 2026 study of commercial systems found that every one exceeded 51 percent word error rate on severe dysarthria ⁷. If your users include people with atypical speech, a benchmark average tells you little, and you need recordings from those users. Errors of a different kind matter too. A 2024 study found that about 1 percent of Whisper transcriptions in its test data contained entire hallucinated phrases or sentences that were not in the audio ⁸. Count invented text separately from ordinary word errors.

What to do: shortlist three providers and test them on your audio

1. Decide the mode first. Recorded files and live speech are tested differently, so choose one before you compare.
2. Shortlist three providers. Include one cloud platform, one speech specialist and one open model you could run yourself.
3. Record 30 minutes of real audio. Include your microphone, your background noise, the names and terms you use, and any speaker changes that matter to you.
4. Write a reference transcript for at least five minutes, and score each provider with the same word error rate rules.
5. Count invented sentences and any speaker labels that are wrong, as separate numbers.
6. Choose the lowest-error option that fits your budget, latency and data terms. If two are close, the cheaper or the safer one usually wins. For price, see the cheapest speech-to-text API.

A provider with the lowest published score is not automatically your best option. Your recordings, your terms and your latency needs decide it.