Modular Full-Duplex Control Module

• The pluggable full-duplex control module is a modular, system-agnostic solution that decouples control logic from domain-specific signal processing for simultaneous transmit and receive operations.
• It leverages FSMs, digital filtering, and LLM-based classifiers to ensure real-time operation, robust error suppression, and precise synchronization across various hardware and software platforms.
• Implemented via in-field upgradable interfaces using digital cancelers and redundancy protocols, it supports applications from RF/optical communications to interactive spoken dialogue systems.

A pluggable full-duplex control module provides a modular, system-agnostic mechanism for real-time, bidirectional flow control in either communication hardware (e.g., RF, optical links) or low-latency spoken dialogue systems. These modules abstract the full-duplex (simultaneous TX/RX) problem by decoupling the control plane from the domain-specific signal or language processing stack. They standardize interfaces for integration, facilitate robust continuous-time or real-time operation, and provide guarantees on error suppression, synchronization, and controllability, often under hardware or environmental constraints. Implementations span high-speed data acquisition, relay self-interference cancellation, and LLM-governed turn-taking for conversational AI.

1. Fundamental Principles and Architectures

Pluggable full-duplex control modules are architected to insert minimal, standardized control logic at the system boundary between hardware front ends (RF, optical, interface boards) or between upstream (audio, ASR) and downstream (NLU, LLM, TTS) units in dialogue systems. In communication hardware, such modules implement digital and analog feedback or filtering between receive and transmit chains, isolating self-interference or providing link-level flow supervision via sampled-data control, protocol transceivers, and protocol-aware redundancy (Sasahara et al., 2015, Huang et al., 2022, Zhang et al., 2018).

Conversely, in dialogue systems, pluggable modules encode the necessary logic for turn management, interruption detection, and barge-in handling atop existing (often legacy, half-duplex) pipelines, employing LLM-based classifiers or signal processing (VAD, EoT) to provide independent, rewritable decision boundaries (Zhang et al., 19 Feb 2025, Chen et al., 8 Sep 2025, Liao et al., 19 Feb 2025). Architecturally, this yields a decoupled, FSM-based or token-based control layer, interfaced by minimal APIs or direct hardware pins, enabling in-field upgrades and subsystem swaps.

2. Mathematical and Control-Theoretical Foundations

In sampled-data full-duplex RF modules, the foundation lies in mixed continuous/discrete H∞ control. The self-interference loop is modeled as a continuous-time, flat-fading, delayed coupling path superposed with receive signal, leading to sampled-data formulations: $$y(s) = a_2P(s)H_h(z)K(z)S_h[a_1re^{-Ls}A_Ly(s) + v(s)]$$ where $K(z)$ is the digital canceler, $e^{-Ls}A_L$ models the delay and RF rotation, and $P(s)$, $S_h$, $H_h$ represent analog and sampling/hold operations. The H∞ synthesis objective seeks

$$\min_{K(z)}\;\|T_{zw}\|_{H∞} = \min_{K(z)}\;\sup_{\omega\in\R}\,\barσ(T_{zw}(j\omega))$$

where $T_{zw}$ is the lower-LFT of the generalized plant with the digital insertion—crucial for optimizing continuous-time attenuation, not simply per-sample energy (Sasahara et al., 2015).

In AI speech systems, dialogue control modules are mathematically realized as LLM-based classifiers (semantic VAD or action classifiers), mapping token/feature buffers to a finite action set—often as argmaxes of cross-attention LLM heads: $c^* = \arg\max_c P(c\mid\text{history})$ or, in FSMs, mapping context, state, and windowed signal features to next-state actions: $\pi_t = F(C, S_{t-1}, W_t; \theta)$ followed by explicit state transitions laid out as matrix or table policies (Zhang et al., 19 Feb 2025, Chen et al., 8 Sep 2025, Liao et al., 19 Feb 2025).

3. Implementation Modalities and Interfaces

Hardware-oriented modules expose power, clock, GPIO, and protocol-specific high-speed interfaces. RF relay cancelers are implemented as digital filters (K(z)), typically IIR (8–12 poles/zeros) or FIR (64–128 taps), updated in-field via SPI/I²C/UART and running on dedicated DSP/FPGAs, with strict timing such that $h < L$ (hold time less than loop delay) to avoid instabilities. Pluggability is defined by the ability to upgrade coefficients, swap out filter cores, or replace front-end modules with standard pinouts (Sasahara et al., 2015, Zhang et al., 2018).

Optical/FPGA-based control modules use SFP+, lpGBT, or VTRx+ transceivers, exposing differential, hot-pluggable lanes (supporting >10Gb/s), with electrical and optical fail-over redundancy—clocks and I²C busses cross-tied such that either module can take over upon fiber or chip failure in <1 ms (Huang et al., 2022). Mechanical pluggability is enforced via standard cages/adapters (SFP+/MPO), tool-less fan-outs, and keyed interconnections ensuring signal mapping and strain relief.

Software dialogue modules present control APIs: process_audio_chunk(), on_state_change(), or VAD.predict(), emitting discrete control signals (pause/resume ASR, TTS; dialogue tokens) while remaining independent of the main model inference stack. This enables plug-and-play replacement, minimal code edits, and rapid module upgrades (Liao et al., 19 Feb 2025, Zhang et al., 19 Feb 2025, Chen et al., 8 Sep 2025).

4. State Machines, Control Protocols, and Decision Rules

Abstracted control flows in these modules are governed by FSMs or token emission policies that encapsulate valid full-duplex dialogue or link-management logic. In FlexDuo, the three-state FSM (Speak, Listen, Idle) is extended with explicit buffering and semantic filtering in the Idle state, with transition matrices defined as: $\delta(S_{t-1},\pi_t) = S_t$ where $\pi_t$ are actions such as “Keep Listening,” “Speak→Listen” (user interrupt), or “Idle→Listen” (speech onset in noise) (Liao et al., 19 Feb 2025). The pluggable module emits only the audio segments and transitions needed by the core SDS, suppressing extraneous cues and minimizing false barge-ins.

In FireRedChat, turn-taking decisions are the result of sequential VAD confidence thresholding, semantic EoT scores, and direct TTS barge-in interrupts. For spoken dialogue LLM modules, the system outputs control tokens—e.g., <|Continue-Listening|>, <|Start-Speaking|>—which are then consumed by the rest of the pipeline to orchestrate stateful, non-blocking full-duplex operation (Zhang et al., 19 Feb 2025, Chen et al., 8 Sep 2025).

5. Robustness, Evaluation, and Redundancy

System-level robustness is a primary design objective. In RF and optical hardware, multipath, temperature swings, and radiation necessitate robust controller synthesis and physical redundancy. Sampled-data H∞ controllers are designed to guarantee closed-loop stability under multiplicative uncertainty (e.g., multipath $\Delta(s)$, bounded as $\|\Delta\|_\infty<1$), with explicit stability and attenuation margins validated by hardware-in-the-loop and over-the-air tests (Sasahara et al., 2015, Huang et al., 2022).

Redundant control (e.g., cross-tied clock/I²C in ATLAS Phase-2 boards) ensures fail-operational capacity without manual switchover. BER (<$10^{-12}$), jitter (<100 ps), and one-way latency (<100 ns) are primary metrics.

Dialogue modules are empirically evaluated on turn-taking F1, false interruption rates, end-to-end latency, and conditional perplexity. FlexDuo, for instance, reduces false interruptions by up to 24.9% against integrated SDS baselines, and explicit Idle states yield significant F1 and false interrupt improvements (Liao et al., 19 Feb 2025). FireRedChat defines system-level barge-in latency (T₉₀), EoT detection accuracy, and overall response latency as key operational metrics (Chen et al., 8 Sep 2025). Lightweight LLM-based VAD controllers maintain >97% correct turn-taking while reducing core LLM compute calls by 90–95% (Zhang et al., 19 Feb 2025).

6. Domain-Specific Applications

Communication Hardware and Data Acquisition

• RF relay cancellation: Pluggable digital self-interference cancelers are critical in single-frequency full-duplex relay stations, enabling simultaneous transmit/receive operation by suppressing looped-back, multi-path induced coupling, with predictable performance in high-gain (a₂~60 dB) scenarios and modular upgradability (Sasahara et al., 2015).
• Optical control in HEP: MicroTCA/AMC boards and ATLAS optical-link controllers implement pluggable duplex architectures for clock, data, and monitoring, ensuring repairability, upgrade paths, and deterministic link-level metrics in hostile environments (kGy, high NIEL) (Zhang et al., 2018, Huang et al., 2022).

Spoken Dialogue Systems

• Dialogue turn management: Modular LLM-VAD and FSM-based controllers provide true full-duplex interaction in conversational AI by mediating turn switches, barge-in interruptions, end-of-turn buffering, and filtering, increasing interaction accuracy and lowering interruption rates (Zhang et al., 19 Feb 2025, Chen et al., 8 Sep 2025, Liao et al., 19 Feb 2025).
• Pluggable upgrades: Modules enable retrofitting legacy half-duplex conversational stacks to full duplex via clean APIs; the main NLU/LLM/TTS pipeline remains untouched, facilitating independent module refinement and deployment.

7. Comparative Overview of Module Types

Domain	Pluggable Module Type	Control Mechanism	Key Metrics / Guarantees
RF/Relay	Digital canceler	Sampled-data H∞ filter K(z)	Loop stability, suppression (dB), BER
Optical/DAQ	SFP+/lpGBT/FMC board	Protocol transceivers, redundancy	Jitter, latency, fail-over time
Dialogue/AI	LLM-VAD/FSM wrapper	Token/FSM decision, semantic VAD	Turn-taking F1, false interruptions

In all cases, pluggability is characterized by standardized interfaces, modular replacement, in-field reconfiguration, and independence from the main processing stack—enabling robust, easily maintainable, and upgradable full-duplex control across disparate physical and algorithmic domains.