NVIDIA Parakeet

About

NVIDIA Parakeet adds offline transcription to TypeWhisper for Windows. It runs the models Parakeet Ultra 0.6B, Parakeet TDT 0.6B and Canary 180M Flash on your PC, so your audio is not sent to a cloud service. On the CPU, the models run through sherpa-onnx. The two Parakeet models can also run on a graphics card from NVIDIA, AMD or Intel through Vulkan, using transcribe.cpp.

Features

• Offline transcription on Windows: audio is processed on your PC
• Three models: Parakeet Ultra 0.6B (recommended), Parakeet TDT 0.6B and Canary 180M Flash
• The Parakeet models cover 25 European languages and detect the spoken language themselves
• Canary transcribes English, German, French and Spanish and can translate them into English
• Processing device: Automatic, CPU or Graphics card. Parakeet runs on graphics cards from NVIDIA, AMD or Intel through Vulkan; Canary always runs on the CPU
• Live preview while you dictate
• Dictionary boosting for Parakeet: the add-on includes the Parakeet CTC component, and you add terms in Dictionary
• Long recordings are split into chunks automatically
• Download, use and remove models on the add-on’s page
• Model files are checked against fixed SHA-256 hashes after the download
• No API key or account required

Transcription Models

Model	ID	Publisher	Size	Languages	Graphics card	Translation
Parakeet Ultra 0.6B	parakeet-ultra-0.6b	Moondream	~670 MB	25	Yes	No
Parakeet TDT 0.6B	parakeet-tdt-0.6b	NVIDIA	~670 MB	25	Yes	No
Canary 180M Flash	canary-180m-flash	NVIDIA	~198 MB	4	No	Into English

The sizes refer to the CPU files. For the graphics card, each Parakeet model downloads a separate version.

Processing Device

Option	What it does
Automatic	Uses the graphics card once its files are downloaded, otherwise the CPU. If the graphics card cannot load the model and the model’s CPU files are present, the model runs on the CPU.
CPU	Works on every PC and runs all three models.
Graphics card	Runs Parakeet on a graphics card from NVIDIA, AMD or Intel through Vulkan. The first switch downloads about 760 MB: the transcribe.cpp runtime and the graphics card version of the model. The add-on uses the first dedicated graphics card, or integrated graphics when there is none.

Setup

1. In TypeWhisper, open Settings > Discover plugins and install NVIDIA Parakeet. It then appears in the Settings sidebar, and its page opens. If you installed it in the setup wizard, open NVIDIA Parakeet in the sidebar instead.
2. Under Models, select Download next to the model you want. Parakeet Ultra 0.6B is the recommended model.
3. Select Use model to use it for dictation.
4. Optional: under Processing device, choose Automatic, CPU or Graphics card.
5. For Canary, set Spoken language in the Dictation settings to English, German, French or Spanish. To translate into English, choose Translate to English as the Transcription task there.

Notes

• Parakeet Ultra 0.6B is Moondream’s retrained version of NVIDIA’s Parakeet TDT 0.6B v3. It has the same size, speed and 25 languages and was more accurate in TypeWhisper’s test: 4.05% against 4.96% word error rate on 50 FLEURS clips in German, English, French, Spanish and Italian. Parakeet TDT 0.6B stays available.
• Canary does not detect the language. If no language it supports is set, it transcribes as English.
• Canary always runs on the CPU, even when Graphics card is selected. If the graphics card fails while you switch, TypeWhisper keeps the previous device.
• Versions before 1.3.0 used NVIDIA CUDA. Version 1.3.0 replaced it with Vulkan: the add-on deletes the old CUDA runtime of about 2.3 GB, and a saved NVIDIA CUDA choice counts as Automatic.
• The add-on stores model files in its plugin data directory under Models/<model-id>/, the graphics card version in the gpu subfolder. Removing a model deletes both.
• Older versions of the add-on were called Local Models (sherpa-onnx).