Hearing It Right, Doing It Safely: A Reflex-Inspired Safety Gatekeeper for Voice-Controlled Exoskeleton Arm Manipulation.

Voice interfaces to anthropomorphic robotic arms are typically evaluated without a formal safety layer, or with safety as a post hoc filter trusting the transcript, leaving transcript corruption unaddressed. This paper presents the DAS3 Neuro-Voice Controller, a voice-driven pipeline for a musculoskeletal arm model, evaluated here as a kinematic surrogate for upper-limb exoskeleton and prosthetic control, built on three commitments: on-device intent classification via a 44 M-parameter DistilBERT classifier over a bounded nine-command vocabulary; a deterministic three-layer Gatekeeper enforcing semantic validation, kinematic feasibility, and hard safety constraints independent of classifier confidence; and an upstream Phonetic Interceptor sanitising characteristic ASR mutilations of anatomical and safety terms ("four arm" → "forearm"). Trained on a class-balanced corpus of 3591 expressions and evaluated on fourteen non-native English speakers, the classifier reached 99.95% top-1 accuracy across 1847 oracle-resolvable utterances; the Gatekeeper reduced Priority Override Failure to 0.0%, rejected 100% of out-of-domain speech, and held false-rejection at 2.02%. Against a keyword-spotter baseline the pipeline delivered a 17.4-fold improvement in end-to-end success; the acoustic front-end remains the main weakness. Upstream input repair and downstream deterministic gating emerge as first-class components of safety-critical voice interfaces for assistive systems.