What is NVIDIA Parakeet, and is it better than Whisper?

NVIDIA Parakeet is a family of open speech recognition models that NVIDIA releases. Parakeet TDT 0.6B v2 is the model people compare with Whisper. It is more accurate than Whisper Large v3 on English benchmarks and much faster ^{1 2}, but it transcribes English only ³. Its successor, Parakeet TDT 0.6B v3, covers 25 European languages ⁴. So Parakeet is better than Whisper for some jobs, and not for all of them.

Our guide to open-source speech recognition lists Parakeet beside the other open families, with their licenses and languages.

Parakeet v2 is English only, and v3 adds 25 European languages

Parakeet TDT 0.6B v2 is a 600-million-parameter model designed for English transcription ³. Parakeet v3 extends support from English to 25 European languages ⁴. Check which version you need before you download anything, because the two are not interchangeable for languages.

On English benchmarks, Parakeet beats Whisper Large v3 and runs far faster

On the Open ASR Leaderboard's English short-form track, Parakeet TDT 0.6B v2 had an average word error rate (WER) of 6.05 percent ¹. Whisper Large v3 had 7.44 percent on the same track ². Lower is better, so Parakeet made fewer errors on those test sets.

Speed is the bigger difference. Parakeet's inverse real-time factor was 3,390, and Whisper Large v3's was 146 ^{1 2}. The benchmark averages several English test sets, so your own speed will depend on your computer and the length of your audio.

Whisper and Parakeet both run on NVIDIA hardware, and Qualcomm builds Whisper for its chips

People searching "Whisper AI NVIDIA" usually want to know whether Whisper runs on an NVIDIA graphics card. It does. The openai-whisper command uses an NVIDIA GPU through CUDA when PyTorch can detect one, and falls back to the CPU otherwise ⁵. A Parakeet model is the natural choice on an NVIDIA GPU, but Whisper runs there too.

The Qualcomm search is about running Whisper on a Snapdragon chip. Qualcomm's Hugging Face repository for Whisper Tiny says it holds "pre-exported model files optimized for Qualcomm devices" ⁶. That build is Whisper, not Parakeet. I did not find a Qualcomm build of Parakeet in the sources checked for this answer.

Parakeet is not better than Whisper on every test or language

Whisper was trained on a broader range of languages. Of its training audio, 117,000 hours cover 96 languages other than English ⁷. If you need a language outside what Parakeet supports (English for v2, 25 European languages for v3), Parakeet does not apply.

Results can also reverse 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 ⁸, against 1.5 percent for Whisper Large V3 Turbo ⁹. One episode is not a benchmark, but it shows that a leaderboard lead can disappear on the audio you actually record.

What to do: pick by language, then test on your own recordings

1. If you need English on an NVIDIA GPU, test Parakeet TDT 0.6B v2 against Whisper on your own clips.
2. If you need one of the 25 European languages, test Parakeet TDT 0.6B v3.
3. If you need other languages, keep Whisper, or test Meta's Omnilingual ASR, which the alternatives question covers.
4. Measure both WER and speed on the same 20 to 30 recordings, transcribed by hand first. The real-time factor gives you the speed comparison.
5. Choose the model with the lower error rate on your audio, and keep its license text with your product.

For the other open options, read the best alternatives to Whisper. If you are asking whether Whisper costs anything, whether Whisper AI is free answers that. For the wider subject, see the speech recognition overview.