Are open-source speech recognition models as accurate as paid APIs?

Almost, but not at the very top. On the Open ASR Leaderboard (version of October 17, 2026, read on October 20, 2026), the best open-weight model is Qwen/Qwen3-ASR-1.7B-hf at an average word error rate of 4.94 percent ¹. The best cloud API, OtterAI/speech-v3.1, scores 3.88 percent ². Lower is better, so the paid API leads by a little more than a point of word error rate. Yet the best open model still beats five of the 13 cloud APIs.

Cloud APIs take the first eight places, then open models join

Eight cloud APIs rank ahead of every open-weight model. The eighth, AssemblyAI's universal-3-5-pro, averages 4.87 percent ³. Qwen3-ASR-1.7B comes next, and the open models then run in a tight pack: CohereLabs/cohere-transcribe-03-2026 at 5.21 percent ⁴, bosonai/higgs-audio-v3-stt at 5.22 percent ⁵ and HojoAI/Hojo-ASR-V1 at 5.22 percent ⁶.

OpenAI's paid API, gpt-transcribe, averages 5.49 percent ⁷, so it trails the best open model. OpenAI's own open Whisper large-v3 scores 6.66 percent ⁸, which puts that open model behind the paid API in points.

Showing 8 of 43 rows. See every row in the dataset.

The open group is far more spread out than the cloud group

The leaderboard ranks 56 ¹³ in all, and most of them are open-weight models. They range from the leaders above to very small models: abr-ai/niagara-9m-batch.en averages 15.27 percent ¹⁴, and google/gemma-4-E4B-it, averages 11.38 percent ¹⁵. The cloud rows span a much narrower range, from the leader to meta/muse-voice-transcribe at 5.75 percent ¹⁶. "Open" therefore says little about accuracy: the model, not the access type, decides.

An open model can also win a single test. HojoAI/Hojo-ASR-V1 has a lower GigaSpeech (podcasts and videos) error rate than every cloud API in the data, even though its average is behind nine of them.

The leading open models have different licenses

Open weights does not always mean free for any use. The leaderboard lists each model's license, and the table above shows it:

• apache-2.0 for Qwen3-ASR-1.7B, Cohere's transcribe model, bosonai's higgs-audio-v3-stt, Hojo-ASR-V1 and IBM's granite-speech-4.1-2b.
• cc-by-4.0 for TheStageAI/thewhisper-large-v3-turbo and nvidia/canary-qwen-2.5b, which allows commercial use with credit.
• Non-commercial licenses for some fast models, such as ibm-granite/granite-speech-5.0-470m-turboctc-nc (5.44 percent ¹⁷), which is licensed cc-by-nc-sa-4.0. A product you sell cannot use it without separate terms.

Read the license text before you build on a model. The leaderboard records only the license name.

What the numbers do and do not measure

The scores are word error rates on nine public English test sets, with capitals and punctuation ignored. The leaderboard's own average also includes two private sets that are not published, so its ranking can differ slightly from the averages here. The audio is not dictation.

Open models have a speed figure (RTFx, how many seconds of audio are transcribed per second of compute) measured on the leaderboard's own hardware, shown in the table. Cloud APIs have no speed figure, so speed cannot be compared across the two groups. Neither group's score includes the cost of running it: open weights need your own hardware and effort, while an API charges per use.

What this means for choosing

If you want the lowest error rate and do not mind sending audio to a vendor, a top cloud API is still ahead, but by a modest margin. If you need audio to stay on your device, a leading open model gives up little accuracy on these tests, and some beat paid APIs. Our on-device benchmark tests models on a Mac with our own recordings. For product-by-product comparison, see the speech-to-text API market and open speech recognition models.

Related pages

• The full dataset and the market overview
• Which cloud API is the most accurate
• Speech recognition accuracy across studies
• How benchmarks are built and the speech recognition overview