MorsEar: Toward Generalizable Low-Resource Covert Messaging via Earable based Inertial Sensing

Paper Title

MorsEar: Toward Generalizable Low-Resource Covert Messaging via Earable based Inertial Sensing

Publication Info

• Topic area: Covert, low-resource text entry using earable IMU sensing for accessibility and constrained environments.
• Keywords: Morse code, earables, IMU sensing, covert communication, accessibility, gesture recognition, symbolic input, low-resource, real-time decoding, adaptive grammar.

Background and Problem

• Problem / challenge: Existing silent or covert input systems are limited by reliance on specialized hardware, small vocabularies, or susceptibility to noise and motion. They often fail to provide robust, low-visibility, language-level interaction.
• Significance: The ability to communicate discreetly and reliably in constrained scenarios (e.g., noisy environments, accessibility needs, or emergencies) is critical for privacy, accessibility, and usability.
• Motivation and related work: Prior systems using EMG, ultrasound, or IMU sensors have explored covert communication but are constrained to small vocabularies, socially awkward inputs (e.g., mouthing words), or poor robustness under noise and motion. This paper revisits Morse code as a versatile, low-bandwidth symbolic system and extends it to earable IMU sensing for continuous language-level interaction.

Solution

• Proposed approach: MorsEar, a real-time, IMU-only earable system that maps near-ear gestures into Morse code for covert text entry.
• Novelty:
  1. Development of an ear-native Morse grammar incorporating gestures for dots, dashes, space, delete, and send.
  2. Physics-aware preprocessing and user-adaptive gesture recognition pipeline for robust operation across users and contexts.
  3. On-device decoding with lightweight autocorrect for real-time, privacy-preserving feedback.
  4. End-to-end system validation in realistic scenarios with diverse participants, including accessibility users.
• Procedure and key techniques:
  1. Signal preprocessing: Bias removal, filtering, and feature extraction from IMU data.
  2. Gesture recognition: A self-supervised pretraining and few-shot fine-tuning pipeline for robust classification of gestures.
  3. Tempo-adaptive grammar: Rolling buffers for symbols, characters, and words, with control gestures for editing and sending.
  4. Autocorrect: Lightweight, on-device language model using prefix-trie beam search and Morse-aware error correction.

Results

• Concrete findings:
  • Silent scenario: CER 7.3%, WER 12.5% → 7.8% (with autocorrect), 8.62 WPM.
  • Café scenario: WER 15.2% → 10.3%, 8.10 WPM.
  • Metro scenario: WER 18.6% → 9.0%, 5.20 WPM.
  • End-to-end latency: Median 47 ms, p95 95 ms.
• Advantage over baselines:
  • Outperforms OESense and Sato et al. in gesture recognition (F1: 0.88 vs. 0.76/0.84 in silent, 0.86 vs. 0.77/0.77 in café, 0.77 vs. 0.61/0.59 in metro).
  • Comparable or superior WPM to constrained-input methods like gaze typing (4–12 WPM) and single-switch scanning (2–6 WPM).
• Experiments / evaluation:
  • 24 participants (including 4 accessibility users) evaluated across silent, café, and metro scenarios.
  • Longitudinal 7-day trial showed stable performance and modest learning gains (1-2 WPM increase).
  • Robustness tested across devices (AirPods, OmniBuds, custom prototype) and activities (walking, chewing, face-touching).
• Limitations and future work:
  • Requires initial familiarization with Morse code.
  • Performance variability across users; broader deployment needed to characterize long-term learning.
  • Limited ecological coverage; future studies should include more diverse activities and longer-term use.

Summary

MorsEar introduces a novel, IMU-only earable system for covert text entry using Morse code. By leveraging a physics-aware preprocessing pipeline, user-adaptive gesture recognition, and on-device decoding with autocorrect, it enables robust, low-visibility communication in constrained scenarios. Evaluations with 24 participants across diverse contexts demonstrate its feasibility, achieving CER as low as 7.3% and WPM comparable to other constrained-input methods. While initial familiarization with Morse is required, the system shows promise as a practical, complementary input channel for accessibility and privacy-sensitive use cases. Future work will explore broader ecological deployment and long-term learning dynamics.