Parakeet vs Whisper: What the Benchmarks Really Show
Parakeet vs Whisper comes down to a trade. NVIDIA's Parakeet is more accurate on short English clips (a 6.32% word error rate against 7.44% for Whisper large-v3 on the Open ASR Leaderboard, as of October 2026) and gets through audio far faster, while OpenAI's Whisper covers 99 languages to Parakeet's 25, can translate into English and runs on almost anything. Below we read all three leaderboard tracks, including the two where Whisper ties or edges ahead, and match each model to the job.
Our own Mac dictation app runs Parakeet, so we've been careful to show where Whisper wins.
Parakeet vs Whisper: the short answer
Pick Parakeet for fast, accurate English (or another European language) and Whisper for everything else. Both are open speech-to-text models, and Whisper isn't simply the slower loser: it ties Parakeet on the leaderboard's five-language multilingual track and edges it on long-form once the CORAAL dataset is removed (published results, checked October 2026).
Key difference: Parakeet is a transducer built for efficiency; Whisper is an encoder-decoder built for breadth.
Parakeet vs Whisper at a glance
Specs come from the Parakeet v3 card, the Whisper large-v3 card and OpenAI's repo.
| Aspect | NVIDIA Parakeet | OpenAI Whisper |
|---|---|---|
| Made by | NVIDIA | OpenAI |
| Models compared | Parakeet TDT 0.6B v3 (and v2) | Whisper large-v3 and large-v3-turbo |
| Size | 600M parameters | 1,550M (large), 809M (turbo); six sizes from 39M |
| Design | FastConformer encoder + TDT transducer decoder | Encoder-decoder transformer, 30-second window |
| Languages | 25 European (v3); v2 is English only | 99 |
| Translation into English | No | Yes (large-v3; turbo isn't trained for it) |
| Timestamps | Word, segment and character | Sentence and word |
| English short-form WER* | 6.32% | 7.44% (large-v3), 7.83% (turbo) |
| Training data | 660,000 h pseudo-labelled + 10,000 h human-transcribed | 1M h weakly labelled + 4M h pseudo-labelled (large-v3) |
| Licence | CC-BY-4.0 | MIT (per OpenAI's repo) |
| Common way to run it on a Mac | FluidAudio (Neural Engine) | whisper.cpp (Metal, Core ML) |
*Open ASR Leaderboard published results, checked October 2026. Lower is better.
What is NVIDIA Parakeet?
NVIDIA Parakeet is NVIDIA's family of open speech recognition models. The one most people mean is Parakeet TDT 0.6B v3, which landed on Hugging Face on 14 August 2025.
A transducer built for throughput
Parakeet pairs a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. TDT predicts each token and how long it lasts, so it can skip ahead where a conventional transducer plods through every frame. NVIDIA's TDT paper reports both better accuracy and faster inference than conventional transducers.
Punctuation, capitalisation and word, segment and character timestamps come straight out of the model, per the v3 card.
Parakeet v2 vs v3
Parakeet TDT 0.6B v2 came out on 1 May 2025. It's English only, trained on about 120,000 hours of English, and scores 6.05% on the leaderboard's short-form track (as of October 2026), which makes it slightly more accurate than v3 on short English clips.
v3 added 24 more languages with automatic language detection. It gave up a fraction on short-form (6.32%) and gained on long recordings: 10.72% against v2's 11.18% on the long-form track. Both have 600M parameters, both are CC-BY-4.0, and both take up to 24 minutes of audio per pass.
The small streaming sibling
There's also parakeet_realtime_eou_120m-v1, a 120M-parameter model that streams audio and detects end of utterance. Its card says it's "designed for use in voice AI agent pipelines". It's English only, outputs no punctuation and uses the NVIDIA Open Model License.
It lets an app tell when you've stopped talking, which anyone cut off mid-thought by a voice assistant will appreciate.
What is OpenAI Whisper?
OpenAI's Whisper is an open, multilingual, multitask speech recognition model, with years of tooling built around it, from whisper.cpp to Mac apps such as MacWhisper.
An encoder-decoder that reads 30 seconds at a time
Whisper is a transformer encoder-decoder that works over a sliding 30-second window (OpenAI's repo). The same model handles transcription, translation into English, language identification and voice activity detection.
Whisper large-v3 covers 99 languages. It was trained on 1 million hours of weakly labelled audio plus 4 million hours of pseudo-labelled audio, and OpenAI says it cut errors by 10% to 20% compared with large-v2.
Six sizes, plus turbo
Whisper comes in tiny (39M), base (74M), small (244M), medium (769M) and large (1,550M), with English-only .en versions from tiny up to medium. The sixth is large-v3-turbo at 809M, which cuts the decoder from 32 layers to 4. Its card is refreshingly candid: "way faster, at the expense of a minor quality degradation".
OpenAI's repo puts turbo at roughly 8 times the speed of large on an A100, with large needing about 10 GB of VRAM and turbo about 6 GB.
An ecosystem that runs almost anywhere
whisper.cpp is MIT licensed and runs on macOS (Intel and Apple Silicon), iOS, Android, Linux, Windows, WebAssembly and Raspberry Pi, with Metal and Core ML support on Apple Silicon and integer quantisation for smaller machines.
Derivatives keep pushing it: on the leaderboard's short-form track, as of October 2026, Distil-Whisper large-v3.5 scores 7.21% and the CrisperWhisper fine-tune 6.67%, both better than large-v3 itself.
The two families now share apps. MacWhisper, built on Whisper, added NVIDIA Parakeet in version 13 in June 2025 (English only at launch), so they often sit side by side.
Parakeet vs Whisper: the key differences
Accuracy depends on which track you read
Parakeet wins two of the four ways the leaderboard slices it, and ties a third. Word error rate (WER) is the share of words a model gets wrong; lower is better, and our guide to speech-to-text accuracy covers the detail. The Open ASR Leaderboard compares 86 open-source and proprietary systems across 12 datasets.
Average WER by Open ASR Leaderboard track (lower is better)
| Track | Parakeet v3 | Whisper large-v3 | Whisper turbo |
|---|---|---|---|
| English short-form | 6.32% | 7.44% | 7.83% |
| English long-form | 10.72% | 11.23% | 11.01% |
| Long-form without CORAAL | 9.10% | 8.67% | 8.72% |
| Multilingual (de, fr, it, es, pt) | 4.81% | 4.81% | 5.56% |
Published results, checked October 2026.
Parakeet leads short-form and overall long-form. Take out CORAAL (a corpus of African American Language) and Whisper large-v3 is slightly ahead on long recordings. It also edges Parakeet on Earnings21, 9.71% against 9.96%. On the multilingual track the two are dead level.
Where the short-form gap lives (Parakeet v3 vs Whisper large-v3)
| Dataset | What it is | Parakeet v3 | Whisper large-v3 |
|---|---|---|---|
| LibriSpeech clean | Read audiobooks | 1.92% | 2.01% |
| LibriSpeech other | Harder audiobooks | 3.59% | 3.91% |
| Earnings22 | Earnings calls | 11.19% | 11.29% |
| VoxPopuli | European Parliament speech | 6.09% | 9.54% |
| AMI | Meeting recordings | 11.39% | 15.95% |
On clean read speech they're almost identical. Parakeet's lead comes from messier, conversational audio, and a four-and-a-half point gap on meetings is the kind you notice.
Neither family tops the board, mind. As of October 2026 the short-form leaders are Cohere Transcribe (5.42%), Zoom Scribe v1 (5.47%), IBM Granite 4.0 1B Speech (5.52%) and NVIDIA Canary-Qwen 2.5B (5.63%). "Parakeet vs Whisper" is a question about two popular open families; the single best model is a different search.
Speed: what RTFx does and doesn't tell you
Parakeet is the far more efficient design, but the leaderboard's speed figures won't tell you how long your own laptop takes. They measure RTFx, seconds of audio processed per second of compute, and the leaderboard runs it in batches on a single data-centre GPU. If you're working through a backlog of recordings on a server, that's exactly the number you want. If you're dictating one email, it isn't.
For the record, as of October 2026: Parakeet v3 manages about 3,333 on short-form against 146 for Whisper large-v3 and 200 for turbo, roughly 23 and 17 times faster. Long-form is 1,003 against 69 and 148 (roughly 15 and 7 times), and multilingual 1,719 against 111 and 176.
The Open ASR Leaderboard paper explains the gap: TDT and CTC decoders "offer superior RTFx, making them better suited for long-form and batched processing".
On a Mac, the runtime (CPU, Metal GPU or Neural Engine) changes real speed a lot, which is the next section's point.
Languages and translation
Parakeet v3's 25 languages are 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 (v3 card).
Quality varies within that list. On FLEURS the card shows Italian at 3.00%, Spanish 3.45% and English 4.85%, but Greek at 20.70%, Latvian 22.84% and Slovenian 24.03%.
Whisper large-v3 covers 99 languages, which makes it the only option of the two for Japanese, Mandarin, Korean, Arabic, Hindi and the rest. Quality varies there too.
Translation is Whisper's alone. Large-v3 translates speech into English, though OpenAI's repo notes that "the turbo model is not trained for translation tasks". Parakeet only transcribes.
Hallucinations and silence
Whisper's decoder generates text conditioned on the audio, and over long silences it can make something up. The Careless Whisper study found "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences", more often when recordings had longer non-vocal stretches. That study tested OpenAI's Whisper API in 2023 and didn't test Parakeet at all.
A transducer, by contrast, can output nothing for a stretch of audio as a normal step, so it has less reason to fill a silence. That's reasoning from the design; nobody has published a measured Parakeet rate.
Long recordings
Parakeet v3 handles up to 24 minutes in one pass with full attention, or up to 3 hours with local attention (v3 card). Beyond that, your script has to chunk the audio. Whisper reads 30-second windows and stitches them together, so length isn't a hard limit. Neither model labels speakers: diarisation needs a separate model.
Where they run
On a Mac, which chip does the work matters about as much as which model you pick. On Apple Silicon, whisper.cpp uses Metal, and the whisper.cpp README says its Core ML encoder on the Neural Engine runs more than three times faster than CPU-only. So if a Whisper app feels sluggish, check what it's running on before you blame the model.
Parakeet on Apple devices usually means FluidAudio, an Apache 2.0 Swift SDK that runs it on the Apple Neural Engine. Its README lists apps using it, including VoiceInk and Spokenly, so you may already have Parakeet without having gone looking for it. Off Apple hardware, it mostly runs through NVIDIA's NeMo toolkit on NVIDIA GPUs, as the v3 card shows. Without either, Whisper is the easier route.
Licence
Parakeet v2 and v3 are CC-BY-4.0, so commercial use is fine with attribution. The EOU 120M streaming model uses the NVIDIA Open Model License, so check those terms separately. Whisper is MIT per OpenAI's repo (the Hugging Face card metadata says Apache 2.0); either way, permissive.
When to use Parakeet vs Whisper
Use Parakeet when
- you dictate or transcribe in English or one of its 24 other European languages
- you want the more efficient architecture (push-to-talk dictation, voice commands) or you're batch-processing lots of audio
- your audio sounds more like a meeting room than an audiobook
- you want a matching small streaming model (EOU 120M, English only) to spot when the speaker stops
Use Whisper when
- you need Japanese, Arabic, Hindi or any other language outside Parakeet's 25, or translation into English (large-v3, not turbo)
- your machine is a Windows PC, a Linux box, an Intel Mac, an Android phone, a browser tab or a Raspberry Pi, or you need a tiny model for weak hardware
- you mostly transcribe long earnings-call style recordings, where large-v3 edges Parakeet (9.71% against 9.96% on Earnings21, and 8.67% against 9.10% on long-form without CORAAL)
Whisper's ecosystem of fine-tunes, distilled models and tooling is a reason on its own if you already rely on it.
Either way, test on your own audio
Your accent, mic and vocabulary can matter more than a one-point leaderboard gap, so run five minutes of your own audio through both.
Why we run Parakeet in Blazing Transcribe
This is ours, so weigh that accordingly. Blazing Transcribe is a Mac dictation app: hold fn, talk, let go, and the text appears wherever your cursor is. By default it runs Parakeet TDT v3 on the Apple Neural Engine through the open-source FluidAudio library. An optional streaming mode switches to NVIDIA's small Parakeet EOU 120M model, which transcribes as you talk and spots when you've stopped. Your voice never leaves your Mac.
We picked Parakeet for the public reasons above: English accuracy, especially on conversational audio; the transducer's throughput advantage; punctuation and capitalisation straight out of the model; and FluidAudio's Apache 2.0 path onto the Neural Engine. In the app, transcription runs 155× faster than real time.
The limits are real. It's Mac only, on Apple Silicon (M1 or newer) with macOS 15 or later. It types dictation into whatever app you're using; it doesn't transcribe files or meetings, label speakers or translate. If you need a language outside Parakeet's 25, translation or Windows, a Whisper-based tool is the better fit.
If you're on a Mac, Blazing Transcribe is free to try.
FAQ
Is Parakeet better than Whisper?
For short English clips, yes: Parakeet v3 scores 6.32% against 7.44% for Whisper large-v3 on the Open ASR Leaderboard, as of October 2026. Elsewhere it's not a clean win. On the five-language multilingual track (German, French, Italian, Spanish and Portuguese) the two tie at 4.81%, and once CORAAL is removed from long-form, Whisper large-v3 edges ahead, 8.67% against 9.10%. Whisper is also the better pick for languages outside Parakeet's 25, translation into English, or hardware that is neither Apple nor NVIDIA.
Is Parakeet faster than Whisper on a Mac?
Its design is far more efficient, with between 17 and 23 times Whisper's short-form leaderboard throughput, but that's batched data from a single data-centre GPU. On a Mac it depends on the runtime: the Neural Engine via FluidAudio, or Metal and Core ML via whisper.cpp.
Should I use Parakeet v2 or v3?
Use v2 for English-only short clips, where it scores 6.05% against v3's 6.32%. Use v3 for any of the other 24 European languages and for long recordings, where it scores 10.72% against v2's 11.18% (published results, checked October 2026).
Can Parakeet translate speech like Whisper?
No. Parakeet transcribes only. Whisper large-v3 can translate speech into English, but OpenAI's repo says turbo isn't trained for translation.
Why do sites quote different numbers for Parakeet and Whisper?
They're reading different things: different tracks, a model card instead of the leaderboard file (the v3 card says 6.34%, the file 6.32%), older snapshots or own-hardware tests. This post uses the leaderboard's published results, checked October 2026.
Does Parakeet hallucinate less than Whisper?
Nobody has published a head-to-head Parakeet vs Whisper hallucination study. Whisper's silence hallucinations are documented, and Parakeet's transducer design gives it less reason to fill silence, but that's design reasoning rather than a measured rate.