Added in v0.17.0
Parakeet speech recognition on Linux
Parakeet is an optional local speech engine in Vocalinux. It runs NVIDIA's Parakeet TDT 0.6B models on your CPU through sherpa-onnx and types what you say into the focused app, with no cloud transcription. It suits people who dictate in English or a European language and want a CPU engine with a short list of models to choose from.
On-device
After the one-time model download, recognition runs locally. Audio stays on your machine.
25 European languages
The default v3-european model covers 25 European languages. v2-english is English only.
Verified downloads
Models come from pinned Hugging Face revisions and are checked against SHA-256 pins before loading.
What is Parakeet?
Parakeet is a family of automatic speech recognition models from NVIDIA's NeMo toolkit. Vocalinux uses the TDT 0.6B models, which are transducer models: they turn audio straight into text tokens rather than generating text the way Whisper does. Vocalinux loads int8 ONNX exports of these models with sherpa-onnx, a speech runtime that runs on an ordinary CPU.
Parakeet arrived in Vocalinux v0.17.0 alongside Faster Whisper. whisper.cpp is still the default engine; Parakeet is something you opt into.
Install and select Parakeet
curl -fsSL https://raw.githubusercontent.com/VocaHQ/vocalinux/main/install.sh -o /tmp/vl.shbash /tmp/vl.sh --auto --engine=parakeet- 1
Install the Parakeet engine
Parakeet is an optional extra. Download the installer and run it with --engine=parakeet. The interactive engine menu does not list Parakeet, so pass the flag. The installer adds sherpa-onnx and points your config at the Parakeet engine.
- 2
Open Settings → Speech Model
Open Settings from the tray icon and go to Speech Model. Expand Advanced to see the Engine and Model size pickers.
- 3
Pick Parakeet and a model
Set Engine to Parakeet, then choose v3-european, v2-english or orukeet-v0.1.0 under Model size. Picking only selects the model; nothing downloads yet.
- 4
Confirm the download
If the model is not on disk, the info card shows its size and a Download button. The transfer starts only after you confirm it. Each file is checked against a pinned SHA-256 before Parakeet loads it.
- 5
Dictate
Hold your dictation shortcut (Right Alt by default on new installs), speak, and release. Text is typed into the focused app.
To start once with Parakeet from a terminal, run vocalinux --engine parakeet. In ~/.config/vocalinux/config.json the engine name is parakeet.
Parakeet models and Orukeet
The Model size picker lists three models for the Parakeet engine. Sizes are the catalog values Vocalinux shows before a download.
| Model ID | Model | Languages | Download | Notes |
|---|---|---|---|---|
| v3-european | Parakeet TDT 0.6B v3 (int8) | 25 European languages | about 639 MB | Default and recommended Parakeet model. |
| v2-english | Parakeet TDT 0.6B v2 (int8) | English only | about 630 MB | Pick this if you only dictate in English. |
| orukeet-v0.1.0 | Orukeet v0.1.0 (int8) | 25 European languages | about 641 MiB | Optional, added in v0.18.0. Weights licensed CC BY-SA 4.0. |
About the optional Orukeet model
Orukeet v0.1.0 was added in v0.18.0. It uses the same local sherpa-onnx recognizer and covers 25 European languages; the recommended Parakeet model and the default whisper.cpp engine are unchanged. Vocalinux downloads it from a pinned revision of oruk/orukeet on Hugging Face, checks every file against bundled SHA-256 pins, and checks the publisher's release manifest against those same pins. Its CC BY-SA 4.0 weights license and NVIDIA attribution are saved with the model.
Languages
With Parakeet, the model decides the language, not the Language picker. Vocalinux treats the language as auto for this engine, so a language you chose for whisper.cpp is not passed to Parakeet. Choose v2-english if you only speak English, or v3-european (or Orukeet) for the 25-language European set.
Two related features do not apply to Parakeet: Follow keyboard layout works with whisper.cpp, Whisper, Faster Whisper and the Remote API only, and custom dictionary terms bias only the Whisper-family engines. Transcript corrections from the custom dictionary do run after Parakeet. See languages by engine, multilingual dictation and custom dictionary.
Hardware and packages
Parakeet runs on the CPU. Vocalinux's Vulkan and CUDA acceleration applies to whisper.cpp, so a GPU does not speed up Parakeet (see GPU acceleration). Plan for roughly 640 MB of disk per Parakeet model, stored under ~/.local/share/vocalinux/models/.
Which packages include Parakeet
| Install method | Parakeet | Notes |
|---|---|---|
| install.sh / source install | Yes | Use --engine=parakeet |
| AUR (vocalinux) | Yes | Listed as an optional dependency: python-sherpa-onnx |
| AppImage (and AUR vocalinux-bin) | No | Ships whisper.cpp only |
| Flatpak | No | Ships whisper.cpp only |
| Snap | No | Ships whisper.cpp only |
| .deb / .rpm | No | Thin packages; whisper.cpp on CPU |
| Nix flake | No | Parakeet extra is not packaged |
Parakeet vs whisper.cpp
Both are local engines. The differences are about languages, hardware and packaging rather than a single winner. For every engine side by side, see the engine comparison; for whisper.cpp sizes, see the Whisper model guide.
| Topic | Parakeet | whisper.cpp |
|---|---|---|
| Runtime | sherpa-onnx running NVIDIA NeMo transducer models | whisper.cpp through pywhispercpp |
| Hardware | CPU | Vulkan (AMD, Intel, NVIDIA), CUDA, or CPU |
| Languages | 25 European languages (v3) or English (v2) | Full language catalog plus Auto-detect |
| Language picker | Ignored; language is treated as auto | Used, including Follow keyboard layout |
| Model choice | Three models of roughly the same size | Tiny to Large v3, with quantized and .en variants |
| Custom dictionary terms | Not applied (corrections still work) | Terms bias recognition; corrections work |
| Packages | install.sh, source, AUR optional dependency | Every package |
Pick Parakeet if you dictate in English or a supported European language on a CPU-only machine and installed with install.sh, from source or from the AUR. Stay on whisper.cpp if you use a language outside that set, want GPU acceleration, rely on custom dictionary terms, or installed the AppImage, Flatpak, Snap, .deb or .rpm.
Troubleshooting
Parakeet is missing from the Engine list
Settings lists only engines whose libraries are installed. Parakeet appears once sherpa-onnx can be imported. Re-run the installer with --engine=parakeet. Installs of this engine through install.sh were broken before v0.18.4, so use a current installer; on Arch, install python-sherpa-onnx.
The model will not load after an interrupted download
Vocalinux refuses to load Parakeet files that fail checksum verification. In Settings → Speech Model, delete the leftover model from Unused downloads and download it again.
Text comes out in the wrong language
The Language picker does not steer Parakeet. If you only dictate English, switch the model to v2-english. For a language outside the European set, use whisper.cpp.
Anything else
Run vocalinux --debug and check ~/.local/share/vocalinux/vocalinux.log, then see the troubleshooting guide.
Parakeet FAQ
Is Parakeet better than whisper.cpp?
It depends on your languages and hardware. Parakeet is a strong CPU option if you dictate in English or one of the 25 European languages its v3 model covers. whisper.cpp stays the default because it covers the whole language catalog, can use a Vulkan or CUDA GPU, and ships in every package. Try both on your own voice and machine.
Does Parakeet send my audio anywhere?
No. Parakeet is a local engine. After the model download, recognition runs on your machine and audio is not sent to a server. Only the Remote API engine and VocaGateway send audio to an endpoint you configure.
Why is the Language picker greyed out or ignored with Parakeet?
Parakeet's language coverage comes from the model, not from the Vocalinux language catalog. Vocalinux treats the language as auto for Parakeet, so pick v2-english for English only or v3-european for the European set. Follow keyboard layout is not available for Parakeet.
Can I use Parakeet from the AppImage, Flatpak or Snap?
No. Those packages ship whisper.cpp only. Use the install.sh installer or a source install with --engine=parakeet, or the AUR package with python-sherpa-onnx installed.
What is Orukeet?
Orukeet v0.1.0 is an optional model for the Parakeet engine, added in v0.18.0. It runs on the same local sherpa-onnx recognizer, covers 25 European languages, takes about 641 MiB, and its weights are licensed CC BY-SA 4.0. Vocalinux downloads it from a pinned Hugging Face revision and checks every file against bundled SHA-256 pins.
Try Parakeet on your machine
Install Vocalinux with the Parakeet engine, or start with whisper.cpp and compare. Also see Faster Whisper on Linux, the other CPU engine added in v0.17.0.