Whisper vs Parakeet: Which On-Device Speech Model Is Better?

Whisper vs Parakeet is the choice most people face when running a speech model on their own hardware, and the two are built on different ideas. This article explains those ideas in plain words, then compares speed, languages, accuracy, timestamps and hallucination behaviour, and ends with a short rule for which to pick. If you only want the rule: Parakeet for English and major European languages when speed and faithfulness matter, Whisper for everything else.

Where they come from

Whisper was released by OpenAI in 2022, trained on a very large amount of audio collected from the web, and published as a family of sizes from tiny to large. The current large model is large-v3, with a large-v3-turbo variant that keeps the encoder and shrinks the decoder for speed. It transcribes around a hundred languages and can also translate speech into English.

Parakeet is a family of models from NVIDIA's NeMo team. The one people mean when they say "Parakeet" is Parakeet TDT 0.6B, a 600-million-parameter model. Version 2 is English-only and became known for sitting at or near the top of the Hugging Face Open ASR Leaderboard while being a fraction of Whisper large's size. Version 3, released in 2025, extends it to 25 European languages. Both are published under a permissive license.

You can run both on a Mac; the Whisper walkthrough covers the Terminal route.

Architecture in plain words

Whisper: listen, then write

Whisper is an encoder-decoder transformer. The encoder listens to a 30-second window of audio and turns it into a compact representation. The decoder then writes the transcript one token at a time, and each new token is chosen by looking at the audio representation and at everything it has already written. That second part matters: the decoder is, in effect, a language model that happens to be conditioned on audio.

The upside is fluency. Whisper produces well-punctuated, well-cased text, copes with accents and background noise because of the sheer variety of its training data, and can use context from the previous window to keep names and spellings consistent. The downside is that writing one token at a time is slow, and a model that is partly a language model will sometimes keep writing when the audio has stopped saying anything.

Parakeet: decide at every moment

Parakeet TDT uses a FastConformer encoder feeding a transducer decoder. A transducer walks through the audio frame by frame and at each step decides: emit a piece of text, or emit a "blank" meaning nothing new was said here. The text is therefore tied to the audio timeline; the model cannot produce a word without an audio frame to attach it to.

The TDT part stands for Token-and-Duration Transducer. A plain transducer must look at every frame, most of which are blanks. TDT predicts, along with each token, how many frames to skip before the next decision. It jumps across silence and across the middle of long vowels instead of inspecting them, which is where much of its speed comes from.

The result is a model that is fast, gives accurate word timings for free, and rarely invents. What it gives up is the language-model fluency: it has less ability to use long-range context, and it is trained on a narrower set of languages.

Speed

This is the least contested difference. Parakeet TDT transcribes many times faster than Whisper at equal accuracy. On a data-centre GPU the published throughput is in the thousands of times real time; on an Apple Silicon Mac running through Core ML, an hour of audio finishes in well under a minute on recent chips. Whisper large-v3 on the same Mac takes several minutes to tens of minutes depending on the chip and whether the Neural Engine is used; small and turbo narrow the gap but do not close it.

If you transcribe a lot of audio, this alone can settle the question.

Languages

Whisper covers about a hundred languages, with accuracy that drops steadily from English through the major European and Asian languages to those with little training data. It also detects the language automatically and can translate any of them into English, a feature Parakeet does not have.

Parakeet TDT 0.6B v2 is English only. Version 3 covers 25 European languages, including the big ones (German, French, Spanish, Italian, Portuguese, Dutch, Polish, Russian) and many smaller ones, with automatic language detection among them. For Japanese, Chinese, Korean, Arabic, Hindi, and most of the rest of the world, Whisper is the only option of the two.

Accuracy

On English benchmarks the two are close, and Parakeet often wins. As of writing, the Open ASR Leaderboard lists Parakeet TDT 0.6B v2 with an average word error rate in the same range as Whisper large-v3 or better, at a fifth of the parameter count. Check the leaderboard for current numbers, since both projects release updates.

Benchmarks are clean, though. In practice:

• Heavy accents and unusual vocabulary tend to favour Whisper's larger models, because they have simply heard more variety.
• Numbers, dates, and formatting favour Parakeet, which was trained to write them the way people type them.
• Long recordings favour Parakeet for a subtle reason covered below.
• Noisy phone audio and overlapping speakers are hard for both. Neither separates speakers; that is a separate step called diarization, explained in speaker diarization explained.

The honest answer is to test both on ten minutes of your own audio. The differences that matter to you will be obvious.

Hallucination behaviour

Whisper's best-known flaw follows from its design. Because the decoder is a language model, feeding it silence, music or noise sometimes produces confident, fluent text that nobody said: a sentence repeated a dozen times, a phrase like "thanks for watching" at the end of a file, or a plausible paragraph over a long pause. It happens most at the start and end of recordings and around gaps. Mitigations exist, such as voice-activity detection to skip silence, disabling the carry-over of previous text, and retrying with a different sampling temperature, and the whisper.cpp and WhisperKit projects ship some of them. They reduce the problem; they do not remove it.

Parakeet's transducer emits a blank when it hears nothing. It has no mechanism for writing sentences that are not anchored to audio, so silence produces silence. Its failure mode is different: a mumbled word comes out wrong or gets dropped, rather than replaced by an invented one. For transcripts someone will rely on without listening back, a model that omits is safer than a model that invents. Very rarely a transducer can still get stuck repeating a token on pathological input, so no model should be trusted blindly.

Timestamps and punctuation

Whisper predicts timestamps as special tokens in its output. They are good enough for subtitles at the sentence level but were never trained to be precise, and word-level timing needs an alignment step on top. Parakeet's timings fall out of the transducer itself: every emitted token is tied to the frame where it was decided, so word timestamps are accurate without extra work. That is why apps that let you click a word and hear the audio at that point tend to prefer transducer models.

Both produce punctuation and capitalisation; Whisper's reads slightly more naturally in long sentences, Parakeet's is consistent.

Long recordings

Whisper processes audio in 30-second windows and, by default, passes the text of the previous window in as context. That keeps spellings consistent, but it also means one bad window can poison the next: a hallucinated sentence becomes context, the model continues it, and the error propagates. Tools that run Whisper often turn the carry-over off for this reason, at some cost to consistency.

Parakeet has no window in the same sense; recent versions use local attention so a long file can be processed in one pass, and it is commonly chunked into pieces of a few minutes anyway. There is no text context between chunks, so nothing propagates.

Running them on a Mac

Whisper runs through whisper.cpp (GPU via Metal, optionally the Neural Engine via Core ML) or WhisperKit (Core ML, Neural Engine by default). Parakeet runs through NVIDIA's NeMo toolkit with PyTorch, or through community Core ML conversions, which is what Mac apps typically embed. To run Parakeet locally on Apple Silicon yourself, the Core ML route is the practical one. Both fit comfortably in memory on an Apple Silicon Mac and both are free to download.

When to pick which

Pick Parakeet TDT when:

• the audio is English or one of the 25 European languages in v3,
• you want the transcript fast, or on a laptop battery,
• you need accurate word timings,
• the transcript will be trusted without listening back, so invented text is worse than a missed word.

Pick Whisper when:

• the language is outside Parakeet's list,
• you need translation into English,
• the audio has heavy accents or rare vocabulary and you can afford the large model,
• you want the most natural-reading punctuation for publication.

Many people end up with both and choose per recording. For local models versus cloud services in the first place, see on-device vs cloud transcription.

Frequently asked questions

Is Parakeet better than Whisper?

For English and the European languages it supports, Parakeet TDT is usually as accurate as Whisper's largest model, several times faster, and much less likely to invent text over silence. Whisper is better when the language is not on Parakeet's list, when you need translation to English, or on heavily accented audio where the large model's breadth helps. Test both on your own recordings before deciding.

Does Parakeet support languages other than English?

Version 2 of Parakeet TDT 0.6B is English only. Version 3 adds 25 European languages with automatic detection, including German, French, Spanish, Italian, Portuguese, Dutch and Polish. It does not cover Asian, Middle Eastern or African languages; for those, Whisper remains the practical on-device choice.

Why does Whisper hallucinate?

Whisper's decoder is trained like a language model, choosing each word from the audio and from the words it has already written. When the audio contains no speech, the second signal dominates and the model produces fluent text that sounds right but was never said. Silence, music, and the ends of files trigger it most. Voice-activity detection and disabling text carry-over reduce it.

Can I run Parakeet on a Mac?

Yes. NVIDIA publishes the weights, and they run through the NeMo toolkit or through Core ML conversions that use the Neural Engine on Apple Silicon. The Core ML route is how Mac apps embed it and is the fastest option on a laptop. There is no official one-line installer like whisper.cpp yet, so most people use an app that bundles it.