What are the best alternatives to Whisper?

For English dictation, a strong open alternative to Whisper is NVIDIA's Parakeet TDT 0.6B v2. On the Open ASR Leaderboard's English tests it made fewer errors than Whisper Large v3, and it processed audio far faster. For the longest list of languages, test Meta's Omnilingual ASR. For devices too small for Whisper, the option is Vosk. Whisper is still the model to try first when you need many languages.

This answer is part of our guide to open-source speech recognition, which compares the main model families on license, language and hardware.

Parakeet beats Whisper on English benchmarks and runs much faster

Parakeet TDT 0.6B v2 had an average word error rate (WER) of 6.05 percent on the leaderboard's English short-form tests ¹. Whisper Large v3 scored 7.44 percent on the same track ². Lower is better, so Parakeet made fewer errors on that test set. NVIDIA's Canary Qwen 2.5B scored lower still on that track, with 5.63 percent, but ran far slower, with an RTFx of 418 ³.

The speed gap is larger. Parakeet's inverse real-time factor was 3,390 and Whisper Large v3's was 146 ^{1 2}. The accuracy gap is small, so one test set cannot settle which model is better for you. The ranking also changes as new systems are added, so check the current leaderboard before you commit.

Omnilingual ASR covers the most languages

Meta's Omnilingual ASR covers more than 1,600 languages ⁴. That is the widest range of any open model in this comparison. A long language list is not proof of accuracy in each language, so test your own language on your own audio before you rely on it.

Parakeet covers fewer languages than Whisper

Parakeet TDT 0.6B v2 is an English model ⁵. Its successor, v3, extends support to 25 European languages ⁶. If your language is outside those 25, Parakeet is not an alternative for you.

Whisper is the broader model. Of its training audio, 117,000 hours cover 96 languages other than English ⁷. For many-language work, Whisper remains the stronger default.

Vosk suits small devices, but its scores need care

Vosk is an offline toolkit. Its project says it works on lightweight devices such as the Raspberry Pi, Android and iOS ⁸. A typical small Vosk model needs about 300 MB of memory while it runs ⁹, which makes it a candidate for phones and small boards.

Vosk publishes word error rates for its models, but they come from its own test sets. Compare them with leaderboard figures only with care. For more on the toolkit, see what Vosk is.

Faster versions of Whisper are variants of it, not alternatives

faster-whisper reimplements Whisper on the CTranslate2 engine ¹⁰, and its project says it is up to 4 times faster for the same accuracy ¹¹. Distil-Whisper is a distilled model with different weights. Its project says it is 6x faster and 50% smaller than Whisper, within 1% word error rate ¹², so test it on your audio like any other model.

Your languages, license and hardware decide, and one test can reverse a ranking

Choose by four things, in this order: the languages you need, the license, the hardware you have, and accuracy on your own audio. Leaderboard rankings can flip on a single recording. In a 2025 9to5Mac test of one 7-minute-31-second episode, NVIDIA's Parakeet v2 had a WER of 12.3 percent, the highest of the three engines the author tested ¹³, against 1.5 percent for Whisper Large V3 Turbo ¹⁴. One episode is not a benchmark, but it shows that a leaderboard lead can disappear on the audio you actually record.

What to do: shortlist two models and test them on the same recordings

1. Write down your languages. If you need English only, Parakeet v2 is a candidate. If you need one of the 25 European languages, check v3. For a wider range, keep Whisper or test Omnilingual ASR.
2. Check the license on each model card. Some models are licensed for non-commercial use only, such as NVIDIA's Canary 1B ¹⁵, so confirm the terms for the version you pick.
3. Match the model to your hardware. Read each model card's hardware notes before you choose. Vosk is the option for small boards and phones.
4. Record 20 to 30 short clips that match your real use, transcribe them by hand, and run each candidate on the same set.
5. Compare WER and speed. Use the real-time factor for speed, and switch only if the gap holds on your recordings.

If you want dictation rather than a model to build on, Yap, which we make, is an on-device dictation app for Mac and Windows. You can download it here. For the choice between Whisper and a hosted service, read whether Whisper is more accurate than Google Speech-to-Text. The speech recognition overview covers the wider subject.