Manual Models and processing
Local models
The speech and language models Neato Echo runs on your PC, and how to choose between them.
Neato Echo keeps your audio on your device. Two kinds of model run locally: speech-to-text models that turn your voice into words, and a language model that writes summaries, tidies dictation, and powers the agent.
NVIDIA Parakeet and Nemotron (the default engine)
Parakeet is the default. It's fast, accurate, and works well on lower-end hardware.
| Model | Size | Languages |
|---|---|---|
| Parakeet TDT 0.6B v3 (default) | ~600 MB | Multilingual (25 languages) |
| Parakeet Unified EN 0.6B | ~631 MB | English (state-of-the-art accuracy) |
| Nemotron Speech Streaming EN 0.6B | ~632 MB | English, streaming (fastest) |
| Nemotron 3.5 ASR Streaming 0.6B | ~650 MB | 15 languages, auto-detect (streaming) |
Setup
- 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.
- Choose Local, then make sure the provider is NVIDIA Parakeet.
- Pick a model and click Download.
The offline Parakeet 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, with no second decoding pass, and with the live preview enabled the text appears as you speak. If the stream fails, transcription automatically falls back to the standard record-then-transcribe path.
OpenAI Whisper (via whisper.cpp)
An alternative engine with a range of sizes. Larger models are more accurate and slower.
| Model | Size | Speed | Quality |
|---|---|---|---|
| Tiny | ~75 MB | Fastest | Basic |
| Base | ~142 MB | Fast | Good |
| Small | ~466 MB | Moderate | Better |
| Medium | ~1.5 GB | Slow | High |
| Large | ~3 GB | Slowest | Best |
| Turbo | ~1.6 GB | Fast | Good |
Setup
- Open Settings, choose Speech-to-Text under AI Models, then the tab for the mode you're setting up.
- Choose Local, switch the provider to Whisper, then select a model.
- Click Download.
The whisper.cpp binary is bundled with Neato Echo. No Python or additional runtime needed.
GPU acceleration
Local Whisper can run on your GPU for much faster transcription. Turn on Enable GPU in the Whisper section:
- NVIDIA: one-click CUDA runtime download from the GPU card in the model picker
- AMD / Intel: 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), Neato Echo automatically falls back to CPU with an in-app notice, and transcription keeps working.
Which speech model to choose
- Parakeet TDT 0.6B v3 (the default): the best all-round choice for speed, multilingual support, and lower-end hardware
- Nemotron Speech Streaming EN: when you want to see text appear live while you speak, and you dictate in English
- Whisper Medium or Large: when accuracy matters more than speed and you have the hardware for it
Language models (Qwen)
Summaries, dictation cleanup and the agent run on a local language model, downloaded in Settings → Language Models. The recommended model is Qwen 3.5 4B (about 2.5 GB). Smaller Qwen sizes are available for PCs with less memory; they're faster but produce simpler summaries.
Where models live
All downloaded models are stored in %USERPROFILE%\.cache\neato-echo. Remove them 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.
Still need help?
Write to [email protected] with your Windows version, your Neato Echo version (Settings → System), and what you were doing. A person reads every message.