Wearable AI Module Research Log

Wearable AI Module Research Log

Project Mission

Design a modular, composable wearable AI system that:

• Reduces cognitive and physical effort for users
• Integrates multiple sensor modalities through adaptive fusion
• Provides intelligent, real-time feedback without overwhelming the user
• Respects user autonomy and preferences
• Scales across different assistive applications

Design Principle: Reduce Effort

All architectural decisions prioritize user experience:

• Minimal manual configuration
• Intuitive feedback and control
• Predictive assistance (anticipate user needs)
• Graceful degradation (work reasonably even when sensors fail)
• User-centered learning (adapt to individual preferences)

High-Level Sensor-Fusion Architecture

WEARABLE MODULE (Embedded System)

├─ Sensor Fusion Engine

│ ├─ Multi-modal sensor input

│ ├─ Real-time feature extraction

│ └─ Sensor calibration and health

│

├─ Embedded AI (TinyML / Edge Inference)

│ ├─ Intent classification

│ ├─ Adaptive synthesis

│ └─ Personalization

│

└─ Output & Feedback

├─ Audio synthesis

├─ Haptic feedback

└─ Vibration alerts

[Wireless + Cloud API]

SMARTPHONE COMPANION APP

├─ User preferences and personalization

├─ Historical data and learning

├─ Advanced analytics and visualization

└─ Cloud connectivity and updates

Potential Sensors

• Contact/bone-conduction microphone
• MEMS accelerometers (3-axis)
• Gyroscope (for orientation)
• Environmental microphone (context awareness)
• Optional: EMG electrode array

AI Responsibilities

1. Sensor Calibration: Adapt to individual physiology and device placement
2. Intent Classification: Determine user's communicative intent (speak, listen, help, cancel)
3. Real-Time Synthesis: Generate pitch, timing, and amplitude assistance
4. Feedback Optimization: Learn user preferences for feedback type and intensity
5. Anomaly Detection: Identify sensor issues or unusual user behavior

Electrolarynx Companion-Module Concept

Scenario: User has a wearable AI module that:

1. Detects speech intent through neck sensors and EMG
2. Predicts the user's intended pitch and timing
3. Automatically activates and adjusts the electrolarynx
4. Provides real-time prosody assistance
5. Learns from user corrections and preferences

Benefits:

• Reduced manual device control effort
• More natural-sounding speech
• Real-time feedback and assistance
• Adaptivity to user state and context

Challenges:

• Accurate intent detection without full acoustic context
• Latency and computational constraints
• User acceptance and comfort
• Clinical validation if used therapeutically

Smartphone Companion Role

The companion app serves as:

• Control interface: Configuration, preferences, manual overrides
• Visualization: Real-time sensor data, intent classification, feedback
• Learning platform: Personalization history, pattern discovery
• Connectivity: Cloud synchronization, future multi-device coordination
• Health and analytics: Usage patterns, comfort feedback, effectiveness tracking

Engineering Questions

1. What is the latency budget for real-time intent detection and synthesis?
2. How should multi-modal sensors be weighted and fused?
3. What ML models are efficient enough for on-device inference?
4. How do we handle sensor failures and degraded modalities?
5. What power budget is realistic for all-day wearability?
6. How should privacy be maintained (data stays on-device where possible)?
7. What firmware update and personalization mechanisms are needed?

Current Status

Concept exploration phase. This is an architectural sketch; no prototypes or evaluations yet. The goal is to establish a mental model for how modular, sensor-fused AI can support assistive applications.

Next Steps

• Refine sensor selection based on feasibility and cost
• Prototype a minimal viable sensor-fusion pipeline
• Explore ML model options for on-device inference
• Conduct preliminary user research on interaction preferences
• Define power and latency budgets