Parakeet vs Whisper on Apple Silicon: Parakeet was 6× faster on our Mac

On an M4 Pro Mac, NVIDIA’s Parakeet V2 transcribed 8 minutes of English speech in 16 seconds. OpenAI’s Whisper Large V3 took 100 seconds for the same file, and made slightly more mistakes.

That’s the headline, and it holds up for English. It doesn’t hold up for Japanese, Chinese or Korean, where Parakeet simply doesn’t work, or for translation, which Parakeet can’t do. This post covers our own Mac numbers, the public leaderboard numbers, why the two models are so different in speed, and how I’d choose between them.

I build TranscribeX, which ships both models, so I have no reason to favour either. The test setup is at the end if you want to reproduce it.

What Whisper and Parakeet are

Whisper is OpenAI’s open speech recognition model, released in 2022 and updated to Large V3 in November 2023 and Large V3 Turbo in late 2024. It comes in sizes from Tiny (39 million parameters) to Large (1.55 billion), handles 99 languages, and can translate speech into English. The code and weights are MIT-licensed.

Parakeet is a family of speech models from NVIDIA. The two that matter here are parakeet-tdt-0.6b-v2 (May 2025, English only) and parakeet-tdt-0.6b-v3 (August 2025, 25 European languages with automatic language detection). Both have 600 million parameters, output punctuation and capitalization, and are licensed CC-BY-4.0, which allows commercial use.

Parakeet vs Whisper speed and accuracy on an M4 Pro

I ran each model twice on the same 8-minute file through the TranscribeX command line and kept the second run, so model loading from disk cache is included but first-launch download isn’t. The highlighted rows are the two models this post is about.

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.

Two results surprised me. First, Parakeet wasn’t just faster, it was also no less accurate. I expected to trade some accuracy for speed. Second, Apple’s on-device recognizer, built into macOS, matched Parakeet’s error rate on this clean audio and was even faster. On noisier, real-world meeting audio I’d expect the gap to open up again, and I haven’t measured that yet.

Parakeet vs Whisper on the Open ASR Leaderboard

Our test uses one machine and one kind of audio. The Hugging Face Open ASR Leaderboard covers eight English datasets, including meetings, earnings calls and accented speech, so it’s a better guide to accuracy across conditions.

Average WER from the Open ASR Leaderboard paper (Srivastav et al., arXiv 2510.06961, Table 3). AMI figures from each model’s Hugging Face model card. Leaderboard speeds are measured on datacenter GPUs, so they’re left out here; use the Mac table above for speed.

The AMI column is the one I care about most, because AMI is real recorded meetings with people talking over each other. Parakeet’s error rate there is about 30% lower than Whisper’s. If you transcribe English meetings, that’s the number to look at.

The leaderboard paper also notes that V3’s extra languages cost a little English accuracy compared with V2. For English-only work, V2 is the better pick.

Why Parakeet is so much faster than Whisper

The speed gap comes from how each model turns audio into text, not from clever engineering in one app.

Whisper is an encoder-decoder transformer. The encoder processes 30-second windows of audio, then the decoder writes the transcript one token at a time, and each token needs another pass through the decoder. Large V3 has 32 decoder layers, so a long transcript means a lot of sequential work that can’t run in parallel.

Parakeet uses a FastConformer encoder with a TDT (token-and-duration transducer) decoder. The decoder is small, and for each step it predicts both the next token and how many audio frames that token covers, so it can skip ahead instead of checking every frame. Most of the work happens in the encoder, which runs in parallel and suits the GPU and Neural Engine in Apple Silicon.

Whisper Large V3 Turbo attacks the same bottleneck from the other side: OpenAI cut its decoder from 32 layers to 4. That makes it much faster than Large V3 with accuracy close to Large V2, but it’s still decoding token by token, and it can’t translate.

When Whisper is still the better choice

• Asian and many other languages. Parakeet V3’s 25 languages are all European. For Chinese, Japanese, Korean, Hindi, Arabic, Thai or Vietnamese, use Whisper (or Qwen3 ASR, which covers 30 languages and 22 Chinese dialects).
• Translation to English. Whisper Large V3 can transcribe and translate in one pass. Parakeet only transcribes, and Turbo can’t translate either.
• Fine-tuned versions. The community has fine-tuned Whisper for specific languages and accents, such as Swiss German, Catalan or Hebrew. Parakeet has far fewer of these.
• Older Intel Macs. Both run best on Apple Silicon. On an Intel Mac, small Whisper models are the practical option.

And one known Whisper weakness worth mentioning: over long silences or music, Whisper sometimes repeats a phrase or invents text that wasn’t said. Apps work around it by detecting speech first, but it’s one more reason I default to Parakeet for English meetings.

Which model to use for what

Running Parakeet and Whisper on a Mac

If you’re comfortable with Python, you can run Whisper through openai-whisper or faster-whisper, and Parakeet through NVIDIA’s NeMo toolkit, though NeMo is built around NVIDIA GPUs. On a Mac, community ports such as whisper.cpp and Core ML or MLX conversions of Parakeet are the faster route.

In TranscribeX, both are a menu choice. The free version includes Whisper Tiny and Apple’s on-device model with no time limit; Pro ($39 once) adds every Whisper size, Parakeet V2 and V3, Qwen3 ASR and Nemotron. You can run the same file through two models and compare the transcripts side by side.

How we tested, and the limits of this test

• Machine: Mac with an Apple M4 Pro and 24 GB of memory, macOS 27.0.1.
• Audio: 73 clips from the LibriSpeech validation set (the hf-internal-testing/librispeech_asr_dummy set on Hugging Face), joined into one 481-second file with 1,150 reference words.
• Method: each model run twice with tx transcribe <file> --model <id> --language en, second run timed end to end. WER computed after lowercasing and removing punctuation.
• Limits: LibriSpeech is clean read speech, much easier than a meeting. 1,150 words is a small sample, so differences under about one point are noise. Whisper Large V3 Turbo and Parakeet V3 weren’t in this round. I’ll add them, plus a real meeting recording, in the next run.