Dictation glossary

Parakeet

Parakeet is a family of open speech recognition models from NVIDIA designed for fast automatic speech recognition. The weights are published so others can run them locally or in their own systems. Variants trade size and latency against accuracy. Products often pick Parakeet when they want on-device or low-latency transcription without sending every utterance to a hosted API.

In more detail

What is Parakeet?

NVIDIA Parakeet sits in the same broad category as other open ASR models: audio in, text out. The design emphasis is throughput and practical local deployment, including relatively compact checkpoints compared with the largest general models. That makes it a common choice for dictation of short utterances, where a user notices hundreds of milliseconds. Long-form accuracy is a different contest; some apps keep a second, slower engine for paragraphs and use Parakeet for the snappy default.

Open weights are not the same as a turnkey app. You still need a runtime (ONNX, CoreML, and similar), voice activity detection, and a desktop wrapper that owns the microphone. Licensing for a given checkpoint is stated by the publisher — Parakeet releases have used a Creative Commons license — and that is separate from whatever terms a dictation vendor adds. If a site says it uses Parakeet, ask which variant, whether it runs on-device, and how large the first download is.

For writers who speak

Why it matters for dictation

Dictation lives or dies on the pause after you stop talking. A model that is accurate but sluggish trains you to type instead. Parakeet-class engines exist to keep that pause short on ordinary laptops, especially for messages and comments. Knowing the name helps you decode a spec sheet that lists two local engines: one for speed, one for long-form care.

In this product

How WhisperJot handles it

WhisperJot's Jot Local engine — the fast on-device default for short utterances — uses NVIDIA's Parakeet models and downloads about 500 MB on a Mac (about 600 MB on Windows and Linux) on first use. Jot Local Pro is the higher-accuracy local alternative. Both run on your device by default and work offline after that download. Jot Cloud is opt-in.

Questions

Straight answers.

What is NVIDIA Parakeet?

Parakeet is NVIDIA's open family of automatic speech recognition models. The models convert speech to text and are published so developers can run them in their own products. They are often chosen for faster local recognition on shorter audio. Parakeet is a model family, not a consumer dictation app; the hotkey, cleanup, and text insertion still come from whatever software wraps it.

How is Parakeet different from Whisper?

Both are open speech-to-text model families you can run without a hosted API. Whisper is a general multilingual recognizer that many long-form tools use. Parakeet is commonly selected when latency on short utterances is the priority. Exact quality depends on the variant, language, and domain. Dictation apps sometimes ship both and let you pick speed versus long-form accuracy.

Can Parakeet run on-device?

Yes. The open weights are meant to be executed in a local runtime, not only in a data center. On-device use still needs RAM, disk for the checkpoint, and often an accelerator such as a Neural Engine or a GPU. If a product streams your audio to a server, it is not using Parakeet as a local engine even if the same model family exists in the cloud.

One hotkey, any focused app.

Private local transcription by default, with an optional opt-in cloud engine.