NVIDIA's Parakeet TDT 0.6B runs well on an Apple silicon Mac: warm, it transcribes an hour of clean audio in under a minute on recent chips, several times faster than Whisper at similar accuracy on English and major European languages. Getting there yourself means a Python environment or a developer toolchain, a download of about 2.5 GB, audio preparation and some care with long files.
This page explains what Parakeet is, why it is fast, which languages it covers, what a manual setup really involves and where it breaks. If what you are after is fast offline transcripts rather than the model itself, the last part shows the one-step way in ThinkScribe.
If you are deciding between model families rather than setting one up, read Whisper vs Parakeet first.
What Parakeet TDT is
Parakeet is NVIDIA's family of open speech recognition models, trained and released through its NeMo toolkit. The two models people usually mean are:
- Parakeet TDT 0.6B v2: English only, 600 million parameters.
- Parakeet TDT 0.6B v3: 25 European languages, same size, and it detects the language of the audio itself.
Both use a FastConformer encoder and a TDT decoder. TDT stands for token-and-duration transducer: instead of deciding something at every audio frame, the decoder also predicts how many frames to skip before its next decision. It jumps across silence and the middle of long vowels, which is a large part of why it is so quick. Output includes punctuation and capitalization, and both models produce word and segment timestamps. Input is 16 kHz mono audio.
The weights are released under CC-BY-4.0, which permits commercial use with attribution. On their model cards, v3 reports a mean word error rate of 6.32% and v2 6.05% on the Hugging Face Open ASR Leaderboard, with throughput above 3,000 times real time. Those speed figures come from data-centre GPUs with large batches, and the leaderboard has since re-scored its test sets (the 2026 open model survey has the current numbers). The v3 card also says the model needs at least 2 GB of memory to load and handles about 24 minutes of audio in one pass with full attention, or up to about three hours with local attention.
What running it yourself involves
Parakeet is published as model weights. There are three kinds of runtime people use on a Mac, and each asks for different skills.
A community Python port for the Mac's GPU. A package that runs the model on Apple silicon graphics cores. It is the quickest route for someone comfortable in Terminal: a Python environment, an audio converter, a command per file or folder, and options for output format, chunk length and precision. The project moves quickly, so options and defaults change between releases.
Converted models for app developers. Conversions aimed at people building Swift apps, running on Apple's dedicated machine-learning hardware. The first use compiles the model for your chip, which is a one-time wait of a few minutes, and you need Xcode and a working Swift toolchain to run them at all.
NVIDIA's NeMo toolkit. The reference implementation. It is a heavy install (PyTorch plus a long list of audio and training dependencies), it runs mostly on the CPU on a Mac, and it is much slower and hungrier than the other two. It is the choice when you need behavior identical to NVIDIA's, for example to reproduce published error rates.
Whichever you pick, the common work is the same:
- Audio preparation. Every route wants 16 kHz mono. Some convert for you, others fail on anything else, so many people convert every file first.
- A 2.5 GB download per model, plus a place to keep it and a way to notice when a newer conversion appears.
- Long files. Recordings longer than a few minutes are split into chunks with some overlap and stitched back together.
- Upkeep. Python versions, package updates and macOS upgrades each have a way of breaking a working setup.
- Everything around the text. Speaker labels, a searchable library, summaries and exports are separate projects.
Speed: Parakeet against Whisper
The Open ASR Leaderboard is the common reference. Its RTFx column (seconds of audio per second of compute, higher is faster) has put Parakeet TDT in the thousands, against well under a hundred for Whisper large-v3 on the same hardware. large-v3-turbo narrows that with a smaller decoder, but the gap is still an order of magnitude, and the reason is structural: Whisper's decoder writes one token at a time with attention over the whole audio window, while a TDT decoder skips ahead by predicted durations.
On a Mac the absolute numbers are far lower and depend on the chip, memory and whether the model is already loaded. Rough expectations for an hour of clean audio:
| Setup | An hour of audio takes | Notes |
|---|---|---|
| Parakeet on a recent M-series chip | Under a minute | After model load and any one-time compile |
Whisper large-v3-turbo on the GPU |
A few minutes | See run Whisper locally on a Mac |
| NeMo on the CPU | Considerably longer | Reference behavior, not speed |
Language coverage
Parakeet v2 is English only. The v3 model card lists these 25 languages, detected automatically:
| Language | Code | Language | Code |
|---|---|---|---|
| Bulgarian | bg | Latvian | lv |
| Croatian | hr | Lithuanian | lt |
| Czech | cs | Maltese | mt |
| Danish | da | Polish | pl |
| Dutch | nl | Portuguese | pt |
| English | en | Romanian | ro |
| Estonian | et | Russian | ru |
| Finnish | fi | Slovak | sk |
| French | fr | Slovenian | sl |
| German | de | Spanish | es |
| Greek | el | Swedish | sv |
| Hungarian | hu | Ukrainian | uk |
| Italian | it |
That is the European set plus Russian and Ukrainian. Audio in a language outside the table produces nonsense rather than an error, so a manual setup needs its own check for which recordings can go through it.
Known limitations
- Timestamps vary by tool. The model produces word timing, but not every port passes it through to its output.
- Chunk boundaries. When a long file is split, a word on a boundary can occasionally be dropped or doubled. Longer chunks help if memory allows.
- Memory. The full-precision weights are about 2.5 GB, and loading briefly needs that plus working space. An 8 GB Mac copes if little else is running.
- No vocabulary hint. Unlike Whisper's prompt, the common Parakeet ports have no way to steer the spelling of names and jargon. You fix those afterwards.
- Cold starts. Every run loads a 2.5 GB model. For a batch, transcribe all files in one run rather than one process per file.
- Very long inputs. Transducers rarely invent text over silence, but inputs far longer than the model was trained on can degrade into repeats. Why speech models repeat themselves explains the pattern.
How to get fast offline transcripts in ThinkScribe
If you came for quick, offline transcripts rather than for the model itself, ThinkScribe gets you there without a toolchain. It sets up the speech model for you: nothing to install or convert.
- Install ThinkScribe on your Mac (Apple silicon, macOS 26 or later) or on iPhone and iPad (iOS and iPadOS 26 or later).
- Choose the spoken language: Settings > Language sets the default, and on iPhone the language chip on the Record screen switches it per recording. There are 28 languages.
- On a Mac, import with Import File... (⌘I) or drag files onto the window; several at once is fine. On iPhone, tap Import Audio File or share a file to ThinkScribe from another app.
- Or record: the record button on iPhone, ⌘R on a Mac.
- Read the transcript with speakers separated, and play from any line to check it against the audio.
- Export as text, subtitles (SRT or VTT), JSON, Markdown, PDF, Word or CSV.
Transcription happens on your device and works offline. There is no account, and your recordings and transcripts are never sent to us. For more on each step, see transcribe audio to text on a Mac, on iPhone, or the audio to text use case. The complete voice to text guide covers live text and voice typing too.
Frequently asked questions
Is Parakeet faster than Whisper on a Mac?
Yes, by a wide margin at similar accuracy on English and major European languages. The transducer design skips ahead through the audio instead of writing one token at a time, so on a recent Apple silicon Mac a warm Parakeet setup transcribes an hour of audio in well under a minute, where Whisper large-v3-turbo needs a few minutes.
Does Parakeet support languages other than English?
Version 3 of Parakeet TDT 0.6B handles 25 European languages including Russian and Ukrainian, detected automatically. Version 2 is English only. Neither covers Chinese, Japanese, Korean, Arabic, Hindi or other languages outside that list.
Can I use Parakeet commercially?
NVIDIA's model cards for v2 and v3 state the CC-BY-4.0 license, which allows commercial use as long as you credit the source. Check the license of any port or conversion you use as well, since the code around the weights has its own terms.
How much memory does Parakeet TDT 0.6B need?
The model card asks for at least 2 GB to load, and the full-precision weights are about 2.5 GB. Budget 2 to 3 GB while loading and somewhat less once it runs at reduced precision. It works on an 8 GB Mac; 16 GB leaves comfortable room for long files.
Can I get fast offline transcription on a Mac without setting up a model?
Yes. ThinkScribe sets up the speech model on your Mac for you, so you import a file or record, pick the language, and read the transcript with speakers separated. It works offline, needs no account, and your recordings are never sent to us.
