Neural Prediction Settings
Fine-tune the neural network-based swipe typing predictions for accuracy and performance.
Quick Summary
| What | Description |
|---|---|
| Purpose | Optimize neural predictions |
| Access | Scroll to Neural Prediction section in Settings |
| Key Setting | Beam Width controls accuracy vs speed |
Settings Location
In Settings, scroll to the Neural Prediction section (collapsible). All neural prediction settings are here.
Understanding Neural Predictions
CleverKeys uses an ONNX neural network to predict words from swipe gestures. The model processes your swipe trajectory and outputs probability distributions for each letter position.
Key Settings
Swipe Typing
Master toggle to enable/disable swipe input:
| Setting | Effect |
|---|---|
| On | Swipe typing enabled |
| Off | Swipe gestures disabled |
Swipe on Password Fields
Allow swipe typing in password fields:
| Setting | Effect |
|---|---|
| On | Swipe works in password fields |
| Off | Only tap typing in password fields (default) |
Beam Width
The most important setting for prediction quality. Controls how many parallel word candidates the decoder tracks:
| Width | Effect |
|---|---|
| 3-4 | Faster, may miss less common words |
| 6 | Balanced (default) |
| 8-12 | More thorough search, finds rare words |
| 16-20 | Maximum accuracy, slower |
[!NOTE] Higher beam width = more word candidates explored = better accuracy but slightly slower. Range: 1-20.
Confidence Threshold
Minimum score for a prediction to be shown:
- Lower = more suggestions, some may be weak
- Higher = only confident predictions shown
Max Word Length
Maximum predicted word length in characters (default: 20). Longer words will be truncated.
Full Neural Settings
For advanced tuning (batch processing, greedy search, ONNX threads, beam search parameters, inference tuning, and presets), tap the Full Neural Settings button at the bottom of the Neural Prediction section. This opens a dedicated activity with all neural parameters.
Tips and Tricks
- Accuracy issues: Increase beam width from 6 to 8 or 10
- Slow predictions: Reduce beam width to 4-5
- Missing words: Check that multi-language is configured correctly
- Long words wrong: Increase beam width
[!TIP] Start with the default beam width of 6. Only adjust if you notice specific issues.
All Neural Settings
| Setting | Default | Range/Options |
|---|---|---|
| Swipe Typing | On | On/Off |
| Swipe on Password Fields | Off | On/Off |
| Beam Width | 6 | 1-20 |
| Confidence Threshold | 0.01 | 0.01-0.5 |
| Max Word Length | 20 | 5-50 |
| ONNX Threads | 2 | 1-8 |
Common Questions
Q: Why are predictions slow?
A: Reduce beam width from 6 to 4-5. Also check if multiple languages are enabled, which increases processing.
Q: Why does it suggest wrong words?
A: Try increasing beam width for more thorough search. Also ensure your primary language is set correctly in Multi-Language section.
Q: How do I add words to the dictionary?
A: Type the word and tap it in predictions to add it to your personal dictionary.
Q: Can I reset neural settings?
A: Use Settings > Backup & Restore to reset to defaults.
Technical Details
The neural model is:
- Format: ONNX (Open Neural Network Exchange)
- Architecture: Encoder-decoder transformer
- Input: Normalized swipe coordinates (x, y, time)
- Output: Per-position letter probabilities
- Decoding: Beam search with vocabulary constraint
Related Features
- Swipe Typing - How to swipe type
- Multi-Language - Multi-language predictions