Work

Clockless Spiking Acoustics via 6-Bit Articulatory Coordinates for Token-Free Speech Processing in Neuromorphic Robotics

Year: 2026
Type: preprint
Abstract: Autonomous cyber-physical robots require real-time speech interaction, yet edge deployment remains fundamentally throttled by the power and latency bottlenecks of conventional computing. Modern speech pipelines route audio through cloud-based deep neural networks and arbitrary subword tokenizers (e.g., Byte-Pair Encoding), consuming kilowatts of data-center power, introducing non-deterministic latency (500–1,500 ms), and severing acoustic sound from the physical mass and motor actuation of the robot. Neuromorphic spiking processors, such as Intel Loihi, were engineered to break this von Neumann power wall through asynchronous, event-driven dynamics, yet the neuromorphic field has largely remained in an "application vacuum," attempting to force-feed synchronous frame-based models and text tokens onto event-based silicon. In this paper, we introduce Neurophonesthemic Coding, an embodied, event-driven speech-to-motor control architecture that directly unifies vocal tract articulatory acoustics, Vector Symbolic Architectures (VSA), and asynchronous spiking hardware. Rather than converting speech into arbitrary symbolic text strings, we map acoustic sound waves directly into an immutable 6-bit Boolean coordinate space (2^6 = 64) decoding four fundamental physical invariants: mechanical resistance, kinematic trajectory, spatial boundaries, and vocalic resonance. This coordinate space grounds 1:1 into a developmental topology of 64 sensorimotor primes (the universal communicative primitives shared by human infants prior to syntactic drift, rooted in Anna Wierzbicka's Natural Semantic Metalanguage framework). Using non-commutative circular delay-line permutations, this 64-concept topology compiles into an immutable 512-byte binary ROM kernel (368 bytes for the foundational spoken subset) with zero token collisions. When deployed within the Intel Lava Neuromorphic Framework simulating Intel Loihi 2 spiking silicon under asynchronous virtual time-stepping, our architecture demonstrates: (1) Zero-Power Quiescence, rejecting ambient room noise up to I = 4.8 with 0 false-alarm spikes across 180 neuron-steps while synchronously latching 6-bit Address-Event Representation (AER) coordinate bursts; (2) Clockless Syllable Pacing, using acoustic Bit-0 transitions to pace cadence and separate staccato commands from soothing speech with a >3x warmth ratio without CPU timer interrupts; and (3) Continuous Spiking Attractors, where a 10-compartment competitive network heals 1-bit corrupted acoustic inputs into correct semantic basins with zero competitor leakage across an 18.8% safety barrier. Operating at 1–2 μJ per semantic decision, representing a one-million-fold (10^6x) energy reduction over cloud LLMs, this architecture solves the core latency and power bottleneck by enabling autonomous robots to bypass arbitrary symbolic translation, interpreting the physical concepts inherent in speech acoustics directly for immediate, low-power cyber-physical attunement. Verified Intel Lava simulation suite and 12/12 unit tests available on GitHub: https://github.com/spout-xy/neuro-phonemic-code
Author: Shea, Brook
Language: English
Open Access status: green
OpenAlex ID: W7235623876
Cites:
Cited by:
Related to: 10