Skip to content
TypeWhisper
DE
Back to Add-ons

Local Models (sherpa-onnx)

Bundled

by TypeWhisper

TranscriptionWindows
Windows Source Code

About

Local Models (sherpa-onnx) adds offline transcription to TypeWhisper for Windows using sherpa-onnx. The plugin currently exposes NVIDIA Parakeet for broad local transcription and Canary 180M Flash for a smaller model with translation support.

Features

  • Offline transcription on Windows through sherpa-onnx
  • Downloadable model files from Hugging Face
  • Automatic, CPU, or NVIDIA CUDA acceleration
  • Managed CUDA runtime installation with a safety probe and CPU fallback
  • Download, load, unload, and remove controls for local models
  • Parakeet model with broad language coverage
  • Canary model with English, German, French, and Spanish support
  • Translation support when the Canary model is selected
  • No API key or cloud account required after the model is downloaded

Transcription Models

ModelIDSizeLanguagesTranslation
Parakeet TDT 0.6Bparakeet-tdt-0.6b~670 MB25No
Canary 180M Flashcanary-180m-flash~198 MB4Yes

Configuration

  • Model - Select either Parakeet TDT 0.6B or Canary 180M Flash.
  • Download - Downloads all required ONNX model files and token files.
  • Language - Canary supports en, de, fr, and es.
  • Acceleration - Use automatic selection, CPU, or NVIDIA CUDA. A changed runtime may require reloading the model.

Setup

  1. Open TypeWhisper Settings > Integrations
  2. Select Local Models (sherpa-onnx)
  3. Download and load the model you want to use
  4. Select the sherpa-onnx provider as your transcription engine

Notes

  • Parakeet is the recommended model for general offline transcription.
  • Canary is the option to use when you need translation support.
  • If the CUDA runtime cannot be loaded safely, TypeWhisper reports the reason and falls back to CPU.
  • The plugin stores model files under the plugin data directory in Models/<model-id>/.