Skip to main content
Local processing keeps your audio on your device. OpenWhispr supports two speech recognition engines: whisper.cpp for the Whisper family, and sherpa-onnx for Orukeet, NVIDIA Parakeet & Nemotron, and Cohere Transcribe.

​
Orukeet (via sherpa-onnx)

The recommended local model, and the one a fresh install lands on when you choose the local route. Orukeet is a multilingual Parakeet fine-tune from Oruk that runs on the same sherpa-onnx runtime as Parakeet, so there is nothing extra to install.

​
Setup

  1. Open Settings, choose Speech-to-Text under AI Models, then the tab for the mode you’re setting up — Dictation, Note Recording or Audio Upload
  2. Choose Local — Oruk is the first tab
  3. Download the model
Oruk’s Orukeet setup guide walks through the same steps for Audio Upload with a sample recording you can check your transcript against.

​
OpenAI Whisper (via whisper.cpp)

The classic local engine, with the widest choice of sizes:

​
Setup

  1. Open Settings, choose Speech-to-Text under AI Models, then the tab for the mode you’re setting up — Dictation, Note Recording or Audio Upload. Each keeps its own engine choice.
  2. Choose Local, then select a Whisper model
  3. Click Download — models are stored in ~/.cache/openwhispr/whisper-models/
The whisper.cpp binary is bundled with OpenWhispr. No Python or additional runtime needed.

​
GPU acceleration

Local Whisper can run on your GPU for much faster transcription:
  • macOS — Metal acceleration is built in on Apple Silicon, no setup needed
  • NVIDIA (Windows/Linux) — one-click CUDA runtime download from the GPU card in the model picker
  • AMD / Intel (Windows/Linux) — one-click Vulkan runtime download from the same GPU card, covering Radeon and Arc/integrated GPUs
If the GPU runtime fails to start (unsupported GPU, out of VRAM), OpenWhispr automatically falls back to CPU with an in-app notice — transcription keeps working.

​
NVIDIA Parakeet & Nemotron (via sherpa-onnx)

A faster alternative to Whisper, especially on lower-end hardware.

​
Setup

  1. Open Settings, choose Speech-to-Text under AI Models, then the tab for the mode you’re setting up
  2. Choose Local, then pick the NVIDIA tab
  3. Download the model
Parakeet’s offline models use INT8 quantized ONNX models for efficient CPU inference. The Nemotron models are streaming models: dictation is decoded live over one persistent connection and the text is committed the moment you stop speaking — no second decoding pass — and with the live preview enabled, partial text updates as you speak. If the stream fails, transcription automatically falls back to the standard record-then-transcribe path.

​
Cohere Transcribe (via sherpa-onnx)

The most accurate offline model OpenWhispr ships — 5.42% average word error rate across 14 languages — and the largest, at about 1.7 GB to download and 2.7 GB on disk. Cohere Transcribe has no automatic language detection, so your dictation language is part of the engine’s setup and changing it restarts the engine. Pick it from the Cohere tab under Local.

​
Which to choose

  • Orukeet — the default: multilingual, fast, and accurate on ordinary hardware
  • Cohere Transcribe — the most accurate, if you have the disk space and don’t need automatic language detection
  • Parakeet & Nemotron — fast alternatives; Nemotron streams text as you speak
  • Whisper — the widest range of sizes, with medium and large for difficult audio
On Apple silicon the sherpa-onnx models use an arm64-native runtime, which makes Orukeet and Parakeet roughly three times faster than they were in 1.9.

​
Disk management

Remove downloaded models anytime from Settings → System → Data Management → Model cache, using Clear cache. Open shows you the folder first. Models can be re-downloaded when needed — see where your files live.