Minimal command AAC: Comparing Classic Signal Processing vs AI on Minimal EEG/EMG

Abstract

This thesis presented a minimal-input AAC system that combines EEG and EMG to enable real-time control of a speech interface using four robust gestures: blink (NEXT), finger tap (SELECT), fist hold (SPEAK), and eyebrow raise (FAVOURITES). To extend functionality without extra hardware, the interface further supports double blink (BACK) and double tap (HOME). Signals are acquired with an OpenBCI Cyton-class board and processed in Python via BrainFlow. The detection pipeline implements: (i) blink detection from bipolar EEG using physiological energy thresholds and a double-event window; (ii) tap detection from forearm EMG using a one-shot finite-state machine with min/max pulse width and a double-event window; (iii) fist as a sustained EMG envelope that suppresses taps while active; and (iv) eyebrow activity from frontal RMS that opens a blink-veto window to reduce crosstalk. A lightweight Tkinter UI maps events to a two-level phrase tree and integrates offline TTS. Beyond this rule-based baseline, the thesis specifies and evaluates two supervised AI detectors trained on raw, session-labelled windows to classify gestures directly. The baseline and AI are compared primarily on classification accuracy, F1 score, and single-vs-double gesture confusion across multi-session recordings, while establishing the architectural framework to assess functional metrics like latency and time-to-utterance in future live trials. Results establish a clear framework for assessing where AI complements the engineered pipeline and where well-designed signal processing already suffices.