Get started
Choosing a model
Find a speech engine that fits your hardware, language, and workflow.
On this page
A speech engine turns audio into text. AI cleanup uses a separate text model afterward. Pick the speech engine first, for your computer, your languages, and how you record.
Start with the default#
For your first recording, keep the default: Parakeet Unified on Windows x64 and Linux x86_64, Parakeet MLX on Apple Silicon with macOS 14 or newer, or Local Whisper elsewhere. Parakeet Unified is English only; if you dictate in another language, switch to Parakeet, Local Whisper, or Qwen3-ASR (see below). Updating never changes an engine you already use.
After that, test alternatives on a short recording you know well. Compare the words, processing time, and memory use on your own computer; a larger model doesn't always do better on your recordings.
Compare speech engines#
| Engine | Where it runs | Supported workflows |
|---|---|---|
| Local Whisper | CPU on all platforms; NVIDIA CUDA on Windows x64 and Linux x86_64 | Dictation with preview, uploads, meetings |
| Parakeet | Windows x64 and Linux x86_64 CPU or NVIDIA GPU; Apple Silicon CPU or Metal; Intel Mac CPU from source | Dictation with preview, uploads, meeting chunks |
| Parakeet MLX | Apple Silicon, macOS 14+; Apple GPU or CPU | Dictation with preview, uploads, meeting chunks |
| Parakeet Unified | Windows x64 and Linux x86_64 CPU, or a GPU from any vendor through Vulkan | English dictation with preview, uploads; meetings with TDT-CTC 110M |
| Qwen3-ASR | Windows x64 and Linux x86_64 CPU, or a GPU through Vulkan; Apple Silicon CPU or MPS on macOS 14+ | Dictation and uploads |
| Nemotron Streaming | Windows x64 and Linux x86_64 CPU or NVIDIA GPU; Apple Silicon CPU or Metal; Intel Mac CPU from source | Dictation with preview, uploads, meetings with native preview |
| Moonshine | Windows x64 CPU; Apple Silicon CPU on macOS 15+ | English dictation, uploads, meetings with native preview |
| API | OpenAI or Groq in the cloud, or a speech server you run | Dictation and uploads |
| Remote computer | The paired host's selected engine | Dictation, uploads, meetings with a supported host model |
Parakeet is a strong all-rounder: on English test clips it made about 40% fewer word errors than the Whisper defaults and finished several times sooner on NVIDIA GPUs. It covers 25 European languages; for others, use Local Whisper or Qwen3-ASR.
Parakeet Unified is the English choice when you want readable live text. It runs two models. Parakeet Unified EN 0.6B, the default, streams its preview: with live preview on, words appear about 2 seconds into speaking and follow the model's own stream, short dictations included, while Parakeet's preview re-decodes 3-second windows and shows nothing for a shorter dictation. It also got fewer words wrong than Parakeet in testing. The final text still comes from decoding the saved recording; a dictation of 30 seconds or more pastes the stream's own final text instead, about 0.3 seconds after you stop. Parakeet TDT-CTC 110M is a compact 135 MB model, bundled in the all-in-one installers, that also handles Meeting Mode; its preview re-decodes short windows.
Check language support#
Language choices depend on the engine. Local engines offer English, Russian, Spanish, French, Portuguese, Mandarin, and Auto where supported.
- Moonshine and Parakeet Unified support English only.
- Parakeet and Orukeet omit Mandarin.
- Parakeet MLX detects the language itself and offers only Auto.
- Qwen3-ASR supports all of these.
- Nemotron includes Mandarin in a broader coverage tier; accuracy may vary.
- Auto detects any language the selected model supports.
When you choose another language while an English-only engine is selected, guided setup and Settings offer a direct switch to a multilingual engine. Language selection transcribes speech; it doesn't translate it. A remote computer offers the choices its host engine supports.
Switch between your languages
In Settings → Dictation → Voice model → Languages I dictate in, choose the languages you use. Switch with the tray's Language menu, a language shortcut, the overlay's language chip, or the Omarchy bar. Switching can apply to the dictation in progress and doesn't reload the engine.
Download models and runtimes#
Settings → Models & storage → Downloads shows each model's size, hardware estimate, required components, publisher, and license. Download weights from Speech models and the engine's runtime from Components.
- Local Whisper on NVIDIA: the GPU Acceleration component needs an NVIDIA driver providing CUDA 12 (525+). The CUDA Toolkit isn't needed. OpenWhisper offers Use this GPU when it spots a capable card. macOS runs Whisper on the CPU.
- Parakeet and Nemotron on NVIDIA: NVIDIA Speech GPU needs a Turing-or-newer card. On Linux, older cards such as the GTX 10 series use NVIDIA Speech GPU (Vulkan) instead. The first time either engine needs a runtime, the prompt names your card and offers the GPU runtime, with the CPU runtime as the alternative.
- Parakeet Unified: the transcribe.cpp runtime comes with the Windows and Linux packages (source checkouts download it, about 30 MB). Download Parakeet Unified EN 0.6B (about 730 MB), or use the bundled TDT-CTC 110M on an all-in-one install. With Auto, it uses a dedicated GPU from any vendor through Vulkan, or else the CPU; choosing GPU also accepts an integrated one. A Linux GPU hasn't been validated yet.
- Apple Silicon: with Auto, Parakeet MLX uses the Apple GPU, Parakeet and Nemotron use NVIDIA Speech GPU (Metal) once installed, and Qwen uses PyTorch MPS. Parakeet MLX is about a 2.5 GB download. Moonshine needs macOS 15.
- Qwen3-ASR on Windows and Linux: it shares Parakeet Unified's bundled transcribe.cpp runtime. With Auto, it uses a dedicated GPU through Vulkan, or else the CPU. Download the GGUF weights once: about 850 MB for 0.6B or 2.2 GB for 1.7B. If you used Qwen3-ASR on Windows before 2.8.96, your selection is kept; download its GGUF weights, and you can remove the old Qwen3-ASR PyTorch runtime there.
The Device choices are Auto, GPU, and CPU. GPU means your graphics card, whichever runtime drives it.
Use a cloud or self-hosted speech server#
Choose API as the engine in Settings → Dictation → Voice model, then pick a Server beside the model, there or in the main window:
- OpenAI (the default) or Groq in the cloud. Add the key in Settings → App → API keys; provider charges may apply.
- A server you run: Speaches, LocalAI, vLLM, whisper.cpp, or any OpenAI-compatible address. These come preset to their usual local address; Edit server… changes it. Start whisper.cpp's server with
--inference-path /v1/audio/transcriptions.
To use another computer's hardware instead, see Remote engines. Privacy & offline lists the network behavior of each option.
Add a custom Whisper model#
Choose Add custom models… from Downloads, Dictation → Voice model, or Meeting Mode → Voice & speakers (with A different engine chosen there). Add a local folder, or find a compatible Hugging Face repository and optional subfolder with Find on Hugging Face.
Custom models must be in CTranslate2 Whisper format with model.bin, a valid config.json, and tokenizer.json. Convert original PyTorch weights first. Removing an entry leaves the source files in place.
Get help from the community, or tell us what went wrong.