NVIDIA Parakeet on your Mac: V3, V2 and the live model

Parakeet comes from NVIDIA’s NeMo team. It uses a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. Instead of writing out text one token at a time the way Whisper does, a TDT model predicts each token together with how many audio frames to skip, so it spends very little time on silence and pauses. That design is the main reason Parakeet is so fast.

Parakeet V2 vs Parakeet V3

From the NVIDIA model cards for parakeet-tdt-0.6b-v2 and parakeet-tdt-0.6b-v3, and the Open ASR Leaderboard paper (arXiv 2510.06961). RTFx was measured on datacenter GPUs with batch size 128 and is useful only for comparing models with each other.

Measured on a Mac

On our M4 Pro test machine, Parakeet V2 transcribed 8 minutes of clean English in about 16 seconds, roughly six times faster than Whisper Large V3 end to end, and with a slightly lower error rate (2.8% against 3.4%). The leaderboard’s 20× figure is pure model throughput on a GPU; our time includes loading the model and preparing the audio, which narrows the gap on a single 8-minute file.

Our own test, run October 17, 2026 on a Mac with an Apple M4 Pro chip and 24 GB of memory, macOS 27.0.1. Audio: 73 clean read-speech clips from the LibriSpeech validation set (hf-internal-testing/librispeech_asr_dummy on Hugging Face), joined into one 8-minute file, 1,150 reference words. Each model was run twice through the TranscribeX command line and the second run is shown; times are end to end, including loading the model. Word error rate uses simple normalization (lowercase, punctuation removed). Clean read speech is easier than a real meeting, so expect higher error rates on your own calls, and treat differences under one point as noise on a sample this size. Qwen3 ASR was not timed in this round; Whisper Large V3 Turbo, Parakeet V3 and Nemotron were not part of it.

Parakeet V3 covers Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. It detects the language by itself. The leaderboard paper notes that the extra languages “come at the cost of English transcription accuracy” compared with V2, which shows up as the small WER difference in the table.

So the choice is simple. If everything you record is in English, use V2. If any of it is in one of the other 24 languages, use V3 and skip the step of choosing a language per file.

How fast is Parakeet compared with Whisper?

On the Open ASR Leaderboard, Parakeet V2 reaches an RTFx of about 3,380 and Whisper Large V3 about 146. On the multilingual sets (German, French, Italian, Spanish and Portuguese), Parakeet V3 and Whisper Large V3 have the same average WER of 4.81%, with Parakeet at roughly 15x the throughput. Those figures come from GPU servers processing large batches, so they do not say how many minutes a file takes on a MacBook. They do show the gap is large, and it is why we suggest Parakeet for English and European-language backlogs.

Speed matters most when you transcribe a lot: a week of meetings, a podcast archive, or lecture recordings. Parakeet V2 also handles up to 24 minutes of audio in a single pass on NVIDIA’s reference setup, and V3 up to 3 hours with local attention.

Parakeet for live transcription: the EOU model

NVIDIA also publishes parakeet_realtime_eou_120m-v1, a much smaller streaming model built for voice interfaces. It transcribes English with 80 to 160 ms of latency and emits an end-of-utterance marker when the speaker stops, with a median detection time of 160 ms. TranscribeX uses it as “Parakeet EOU 160ms” for low-latency live transcription.

It has trade-offs. It outputs no punctuation or capitalization, and its average WER is 9.30% at 160 ms, higher than the full Parakeet models. For live captions in other languages or with punctuation, TranscribeX uses NVIDIA Nemotron instead. After a meeting, you can always re-transcribe the recording with Parakeet V2 or V3 for the final text.

EOU figures from the parakeet_realtime_eou_120m-v1 model card (NVIDIA Open Model License).

Parakeet on a Mac

NVIDIA built Parakeet for its own GPUs, and the reference code is NeMo on CUDA. To run on a Mac, the model has to be converted. TranscribeX runs Parakeet locally on Apple hardware, using the Neural Engine on Apple Silicon. In the app, Parakeet V2 is a 476 MB download, small next to Whisper Large V3’s 3.1 GB. You can also pick Parakeet as the model for dictation, where its speed keeps the delay between speaking and seeing text short.

Parakeet is part of TranscribeX Pro, along with speaker diarization, AI summaries and automatic meeting processing. MacWhisper also offers Parakeet; see how the two apps compare.

When Whisper is still the better choice

• Your audio is in a language Parakeet V3 doesn’t cover, such as Japanese, Arabic, Hindi or Turkish. Use Whisper or Qwen3 ASR.
• You want speech translated into English during transcription. Whisper Large V3 can do that; Parakeet only transcribes.
• You switch languages within one recording a lot. Test both on your own audio.