Swipe Typing Technical Specification

Overview

Neural swipe typing uses an ONNX transformer model with beam search to predict words from gesture paths.

Key Components

Component	File	Purpose
Trajectory Processor	SwipeTrajectoryProcessor.kt	Convert touch points to key sequence
Neural Engine	NeuralPredictionEngine.kt	ONNX model inference
Beam Search	BeamSearchDecoder.kt	Find top-k word predictions
Vocabulary	OptimizedVocabulary.kt	Dictionary and trie lookup
Keyboard Grid	KeyboardGrid.kt	Map coordinates to keys

Architecture

Touch Events (Pointers.kt)
    ↓
SwipeTrajectoryProcessor
    ↓ (key sequence)
NeuralPredictionEngine
    ↓ (token probabilities)
BeamSearchDecoder
    ↓ (top-k candidates)
SuggestionHandler → UI

Gesture Sampling Robustness

There is no minimum-speed gate on swipe typing — a slow swipe is not rejected for being slow. Word activation depends on registering ≥2 keys plus swipe_min_distance of path, which a slow-but-complete swipe satisfies just like a fast one (path length is bounded by key geometry, not by speed).

ImprovedSwipeGestureRecognizer.addPoint does, however, drop samples whose inter-sample gap exceeds MAX_POINT_INTERVAL_MS (500 ms). To keep a mid-gesture pause (a deliberate swiper holding still to aim) from permanently stalling the gesture, the recognizer re-anchors _lastPointTime to the resume timestamp when a long gap is seen, then resumes on the next sample. Without this re-anchor a single >500 ms gap left _lastPointTime stale, so every later sample’s delta grew larger and the remainder of the swipe was dropped — the second key never registered and no word was produced.

Neural Model

Property	Value
Format	ONNX Runtime Mobile
Architecture	Transformer encoder
Input	Key token sequence
Output	Probability distribution over vocabulary
Size	~2 MB per language

Beam Search Configuration

From Config.kt:

Setting	Key	Default	Range	Source
Beam Width	neural_beam_width	6	1-32	Config.kt:130, validator backup/SettingsValidation.kt
Max Length	neural_max_length	20	10-50	Config.kt (NEURAL_MAX_LENGTH)
Confidence Threshold	neural_confidence_threshold	0.01	0.0-1.0	Config.kt:132 (NEURAL_CONFIDENCE_THRESHOLD)
ONNX Models	bundled assets	—	—	src/main/assets/models/swipe_{encoder,decoder}_android.onnx

Key Methods

SwipeTrajectoryProcessor.kt

// Line ~120: Convert swipe to tokens
fun processTrajectory(points: List<TouchPoint>): List<Int>

// Line ~180: Get nearest key for point
fun getNearestKeyToken(x: Float, y: Float): Int

NeuralPredictionEngine.kt

// Line ~80: Run inference
suspend fun predict(tokens: IntArray): FloatArray

// Line ~150: Load ONNX model
fun loadModel(context: Context, language: String)

Performance Metrics

Typical inference times:

Device Tier	Inference Time
High-end	15-25 ms
Mid-range	30-50 ms
Low-end	80-150 ms

Related Specifications

• Neural Prediction - Deeper architectural reference (beam search algorithm, token mapping, model I/O shapes, memory pooling)
• Gesture System Overview - Touch event routing and hasLeftStartingKey gatekeeper
• Autocorrect Specification