Parakeet TDT 0.6B v3 vs Whisper Large V3 Turbo

Whisper Large V3 is accurate but heavy: 32 decoder layers, around 1.55 billion parameters, and slow enough that real-time dictation is not its natural home. So OpenAI shipped a distilled build, Whisper Large V3 Turbo, that trims the decoder to 4 layers and drops to around 809 million parameters. Turbo is the one people reach for when they want Whisper-family accuracy at a speed that keeps up with live dictation. That makes it the fair thing to put against Parakeet TDT 0.6B v3, which was built for exactly that job from the start.

The reference for English accuracy is the Hugging Face Open ASR Leaderboard, which scores models on shared English test data and reports an average word error rate. Parakeet TDT 0.6B v3 posts 6.32% there, against Whisper Large V3’s 7.44%. Turbo is a distilled build of Large V3 that trades a little accuracy for speed, so it lands at or just behind Large V3, which puts it behind Parakeet v3. Lower is better, and Parakeet wins the headline, but the honest framing is that the margin is roughly one word in a hundred.

Keep the number in proportion. Word error rate is an aggregate over a fixed English corpus; your own accuracy turns on your accent, your microphone, background noise and how much domain vocabulary you use. The leaderboard fairly ranks models against each other, but it is not a promise about your transcripts. The fuller account of what the figure does and does not tell you is in the primer on how word error rate is actually measured.

Turbo exists to be fast, and it is: dropping the decoder from 32 layers to 4 makes it several times quicker than full Large V3 at close to the same accuracy. But it is still an encoder-decoder model, and it is still slower than Parakeet on the same hardware. NVIDIA’s published benchmarks put Parakeet TDT 0.6B v3 at around a 3,333x real-time factor on Apple Silicon; a ten-minute dictation transcribes in a fraction of a second.

For batch transcription the difference between “fast” and “faster” barely registers. For push-to-talk dictation it dominates how the tool feels, because the whole point is that the words appear the instant you release the key. This is the axis where the choice of Parakeet stops being about a benchmark and starts being about the experience of typing with your voice.

This is the part the leaderboard does not capture. Whisper is an encoder-decoder model whose decoder always wants to produce text; in long pauses or near-silent audio it can hallucinate words. Turbo inherits that behavior, because distilling the decoder makes it smaller, not differently shaped. Parakeet is a Token-and-Duration Transducer: it predicts tokens against the audio stream incrementally and models duration explicitly, so a long pause in the audio becomes a long pause in the transcript rather than an invented sentence.

For dictation this matters more than a one-point word error rate gap, because dictation is full of pauses. You stop to think, you look something up, you trail off mid-sentence. A transducer stays quiet through all of that; an encoder-decoder model is the one that occasionally fills a silence with a phrase you never said. The general architecture explainer is in NVIDIA Parakeet on a Mac.

Turbo keeps Whisper’s roughly hundred-language coverage. Parakeet TDT v3 covers 25 European languages with automatic detection: English, French, German, Spanish, Italian, Portuguese, Dutch, Polish and the rest of the European set. The rule is simple and it is not an accuracy trade, it is a coverage one. Inside those 25 languages, Parakeet is the model to want for its speed and silence behavior. Outside them, in Mandarin, Japanese, Korean, Arabic, Hindi or anything else, Parakeet does not cover you at all, and Turbo is the correct choice. People working across languages can read the role-specific take in dictation for translators.

• You dictate in a European language and want it instant. Parakeet TDT v3: faster on Apple Silicon, slightly ahead on the English benchmark, and it does not hallucinate in your pauses.
• You dictate in a language outside the 25 European ones. Whisper Large V3 Turbo is the only one of the two that covers you.
• You want the highest raw accuracy and speed is no object. Full Whisper Large V3 edges Turbo on accuracy but is far slower; for dictation that trade rarely pays.
• You want the model with no picker to configure. An app built around a single on-device model is the simpler path, which is the bet Parakeety makes.

Parakeet is the NVIDIA model; Whisper is the OpenAI one; Turbo is OpenAI’s distilled build of Whisper Large V3. Parakeety, with the extra letter, is this app: a Mac dictation tool that runs the Parakeet TDT v3 model on-device. It is not affiliated with NVIDIA. The comparison above is between the two models, so it holds for any app running them, not only for ours.

Parakeety runs Parakeet TDT v3 on the Apple Neural Engine, which is why dictation feels instant. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into. Audio never leaves the machine. It needs Apple Silicon and macOS 14 or later. There is a free 7-day trial with no card required. After that it is $30 once.

Try Parakeety free →