Skip to main content

​
Model Selection

Echophrase ships with a small registry of pre-vetted models. The values below mirror the desktop app’s model registry — they are the exact models you can pick from in Settings.

​
Available Models

​
Language coverage

  • Parakeet-TDT 0.6B v3: 25 European languages with automatic detection
  • Qwen3-ASR 0.6B: 52 languages including Chinese (Mandarin, Cantonese, and 20+ dialects), Japanese, Korean, Arabic, Hindi, Thai, Vietnamese, and the major European languages — with strong mixed English/Chinese code-switching in a single sentence
  • Parakeet Lite: English only
Speak Chinese, Japanese, Korean, or another non-European language? Pick Qwen3-ASR 0.6B — it’s the only model in the registry that covers Eastern languages. Echophrase auto-recommends it on non-English systems with a capable GPU, and pairs it with an optional Simplified ↔ Traditional Chinese output toggle (Settings → Voice Model → Chinese script).

​
Model formats

The bundled models are all ONNX. Custom (bring-your-own) models must also be ONNX - see Custom Models for how to import one. GGUF, PyTorch .bin, and safetensors are not supported for custom models in this release.

​
Recommendations

  • CPU only: Parakeet Lite
  • Modern GPU (DirectX 12): Parakeet-TDT 0.6B v3 for best quality, or Lite if you want lower memory use
  • Non-English or multilingual dictation: Qwen3-ASR 0.6B (4 GB+ GPU)
  • Older / integrated GPU: Parakeet Lite

​
Changing Models

  1. Open Echophrase settings
  2. Navigate to Voice Model
  3. Click Change and pick a model
  4. Wait for the download to complete (~1–2 minutes)
Models are downloaded once and cached locally. Switching between downloaded models is instant.
More models on the way. Echophrase is model-agnostic, so new backends can be added without a rewrite. We’re evaluating additional multilingual voice models - see the Requirements page for the latest matrix.