Full-Duplex Interaction: Systems & Applications

• Full-duplex interaction is the simultaneous two-way exchange of information within a single channel, achieved through advanced self-interference cancellation.
• Systems employ a hybrid approach combining passive RF isolation, active analog, and digital baseband cancellation to significantly boost throughput and reduce error rates.
• Applications extend to wireless networks, dialogue systems, and quantum/visible light communications, emphasizing low-latency synchronization and robust interference management.

Full-duplex interaction refers to the simultaneous, bidirectional exchange of information between two or more entities—such that both can transmit and receive at the same time—across a shared communication medium or interface. In engineered systems, it stands in contrast to half-duplex or turn-based interaction, where transmission and reception are mutually exclusive in time or frequency. Full-duplex operation is foundational to modern wireless networks, emerging spoken dialogue systems, and quantum and visible light communications, and requires precise management of self-interference, synchronization, and behavioral arbitration to achieve seamless and natural two-way information flow.

1. Physical Layer Principles and Self-Interference Cancellation

Practical full-duplex physical layer systems are fundamentally constrained by self-interference (SI)—the phenomenon where a device's transmitted signal leaks, often at much higher power, into its own receiver. Overcoming SI is requisite for enabling simultaneous transmit and receive chains in the same spectral band. Modern RF implementations employ a combination of:

• Passive RF isolation through hardware (antenna separation, directional couplers, spatial separation, and circulators).
• Active analog cancellation via multi-tap delay lines, phase shifters, and gain/attenuation elements, which synthesize a replica of the SI channel (as measured, e.g., by a network analyzer) and subtract it in the analog domain before the ADC. The cost function

$$\min \left(y(t) - \sum_{i=1}^M x_{\tau_i,a_i,\phi_i}(t) \right)^2$$

is optimized over tap parameters to cancel leakage (Smida et al., 2023).

• Digital baseband cancellation, where the system estimates the SI channel after analog cancellation and subtracts a digitally synthesized copy in the digital domain using precise estimates of the SI channel impulse response:

$$H_{aa}(f) = \frac{1}{2} H_{+}^{(\mathrm{RF})}(f + f_c), \quad h_{aa} = \mathcal{F}^{-1}\{ H_{aa}(f) \}$$

Adaptive filters (such as sparse RLS with basis expansion) track fast time-varying SI, particularly in underwater acoustic channels (Shen et al., 18 Jan 2024).

This hybrid approach, leveraging measured SI channels and separate analog and digital cancellation—as in the use of four-layer co-integrated patch antennas—enables not only up to 2.4 bps/Hz rate gains over pure RF-isolated configurations but also improvements in BER between 1–4 orders of magnitude (Kaufman et al., 2013, Kaufman et al., 2013). However, active cancellation bandwidth is typically limited; passive plus digital baseband schemes excel in wideband scenarios, important for modern OFDM and high-rate applications.

2. Duplexing in Networks, Scheduling, and Interference Management

In networked settings, especially cellular and Wi-Fi, full-duplex operation introduces complex new interference phenomena: beyond SI, intra-cell and inter-cell interference scale with network density. Achieving near-double spectral efficiency, as theoretically possible for point-to-point links (Li et al., 2016), is contingent on careful resource and interference management.

State-of-the-art strategies include:

• Hybrid Scheduling: Algorithms that opportunistically enable full-duplex communication only when the net utility gain (e.g., in throughput or fairness) outweighs the induced interference cost. This is formalized as a joint optimization over user pairing and transmit power using convex formulations (e.g., geometric programming for power allocation), with utility approximations:

$$\chi^{d}_{(b,\psi^d(b))}(t) \approx w_{(b,\psi^d(b))} \cdot R^{d}_{(b,\psi^d(b))}(t)$$

(Goyal et al., 2014).

• Probabilistic MAC Control: Distributed protocols with epoch-based probabilistic contention and transmission opportunity allocation, derived via LP optimization subject to traffic, fairness, and time constraints, enable adaptive mode switching between full-duplex and half-duplex contingent on traffic balance and interference (e.g., inter-client interference in three-node scenarios) (Chen et al., 2016).
• Interference-Aware Pairing and Binary Power Control: Optimal pairing that minimizes co-channel interference (choosing spatially separated pairs) and binary (on/off) power allocation for each direction, potentially reverting to half-duplex dynamically in high-interference contexts (Li et al., 2016).

Empirical results confirm that, with practical SI cancellation (95–110 dB), full-duplex can nearly double throughput in small-cell indoor deployments (up to 94% improvement), with smaller but still substantial uplink and downlink gains in outdoor, interference-limited scenarios (Goyal et al., 2014). Energy efficiency, however, tends to decrease unless further architectural innovations are introduced (e.g., enabling FD at the UE or penalizing high power in the scheduler).

3. Architectures and Synchronization in Dialogue and Human-AI Interfaces

Full-duplex interaction in speech dialogue systems (FDSDS) departs from classical turn-taking by supporting synchronous speaking, listening, and interruption—mimicking natural human conversation with overlapping speech, backchannels, and rapid turn-taking. Architectures are divided along the Engineered Synchronization and Learned Synchronization axis (Chen et al., 18 Sep 2025):

• Modular Control (Engineered Synchronization): Systems such as FlexDuo (Liao et al., 19 Feb 2025), FireRedChat (Chen et al., 8 Sep 2025), and integrated pipelines with external turn-taking controllers employ finite state machines (FSMs), neural voice activity detection (VAD), and explicit semantic endpoint detectors (e.g., Phoenix-VAD (Wu et al., 24 Sep 2025)). Explicit states (Speak, Listen, Idle) and control tokens arbitrate simultaneous perception and response, often yielding explicit, modular designs that can filter background noise, maintain semantic integrity via buffered input, and allow pluggable upgrades to half-duplex LLMs.
• End-to-End Synchronous Modeling (Learned Synchronization): Models such as Synchronous LLMs (Veluri et al., 23 Sep 2024) extend pretrained autoregressive LLMs (Llama3-8b) to consume and generate interleaved, time-synchronized token streams mapped to fixed-duration real-world clock chunks. Training leverages hundreds of thousands of hours of synthetic and real spoken dialogues to induce temporal behaviors (overlaps, anticipatory responses, backchannel generation) directly into the model’s latent representations.

FSMs in these architectures are often defined via a minimal set of transition tokens (e.g., [C.SPEAK], [S.LISTEN] etc. (Wang et al., 29 May 2024)), enabling real-time decision making at the sub-utterance level and allowing both seamless interruption management and low-latency response.

4. Benchmarking, Evaluation, and Metrics

Recent works establish that traditional turn-based dialogue evaluation is insufficient for full-duplex systems (Chen et al., 18 Sep 2025, Peng et al., 25 Jul 2025, Ge et al., 26 Sep 2025). Comprehensive benchmarking incorporates:

• Behavioral Arbitration: Metrics such as Interruption Response Delay (IRD), Interruption Success Rate (ISR), Success-Interrupt Rate (SIR), Success-Reply Rate (SRR), and Early-Interrupt Rate (EIR) directly assess a system’s ability to handle interruptions, backchannels, and spontaneous user input.
• Temporal Dynamics: First Speech Emit Delay (FSED), First Talk-Over (FTO), and median/long-tail (P₅₀/P₉₅) latencies quantify timing fidelity (e.g., FireRedChat achieves T₉₀ barge-in accuracy at 170 ms and P₅₀ latency at 2.341 s (Chen et al., 8 Sep 2025)).
• Semantic Coherence and Dialogue Quality: Conditional Perplexity (c-PPL) and GPT-4o-rated topic shift and emergency response relevance discriminate models’ contextual accuracy, particularly in high-stakes and interruption-rich scenarios (Ge et al., 26 Sep 2025).
• Acoustic and Backchannel Metrics: Jensen–Shannon Divergence (JSD) for backchannel timing, Naturalness MOS (N-MOS), and acoustic Word Error Rate (WER).

Benchmarking pipelines (e.g., FD-Bench (Peng et al., 25 Jul 2025) and FLEXI (Ge et al., 26 Sep 2025)) with controlled synthetic conversations (including emergencies, user/model interruptions, and backchannels) reveal that even state-of-the-art models face consistent challenges with real-time interruption, semantic coherence under noise, and managing false triggers from acoustic events.

5. Implementation Strategies and Practical Applications

Notable findings in full-duplex interaction system design and deployment include:

• Pluggable Modularization: Systems such as FlexDuo (Liao et al., 19 Feb 2025) and FireRedChat (Chen et al., 8 Sep 2025) enable upgrading of existing SDS or robotic platforms with full-duplex control layers (e.g., explicit Idle state filtering, semantic buffering) without retraining underlying LLMs or core dialogue engines (CDEs). This decoupling supports rapid iteration and domain adaptation.
• Streaming and Low-Latency Processing: Streaming ASR and TTS modules, paired with time-aware FSM-driven LLMs, enable sub-second response latency, with median delays of 0.68 s and more than threefold reduction versus half-duplex systems (Wang et al., 29 May 2024).
• Personalized VAD and Speaker Verification: Integration of target-speaker embeddings (e.g., ECAPA-TDNN) into streaming VAD modules yields robust barge-in control, suppressing spurious interruptions from competing speakers or environmental noise (Chen et al., 8 Sep 2025).
• Semantic Endpoint Detection: LLM-based endpoint detectors (e.g., Phoenix-VAD) use sliding window inference and semantic context to determine when to continue listening or to yield, and achieve F1 scores >0.9 under both semantically complete and incomplete utterances (Wu et al., 24 Sep 2025). Timeouts adapt to utterance completeness, reducing false starts.

Applications span lifelike voice assistants, real-time customer service, on-chip integration (for visible light and quantum full-duplex), networked cellular systems, cognitive radio, and ISAC (integrated sensing and communications), where the joint optimization of communication and sensing tradeoff is solved via SCA-based iterative algorithms and beamforming (Wang et al., 2022).

6. Theoretical Foundations and Future Research Directions

Mathematical formalisms underlying full-duplex interaction have expanded from physical-layer SI channel modeling to neural dialogue joint probability objectives. Joint next-token or next-token-pair prediction (NTPP) is emerging as a unifying paradigm for full-duplex language systems:

$$P(S_e, S_a) = \prod_{t=1}^T P(e_t, a_t | S_e^{(<t)}, S_a^{(<t)})$$

This framework supports simultaneous generation of user and agent tokens, potentially enabling truly parallel dialogue with minimal interaction-induced latency (Chen et al., 18 Sep 2025, Ge et al., 26 Sep 2025).

Advancement in full-duplex interaction is guided by:

• The aggregation and curation of synchronously annotated, multi-channel datasets capturing overlaps, interruptions, and naturalistic timing, needed for effective end-to-end system training (Chen et al., 18 Sep 2025).
• Integration of safety and robustness at ultra-low latencies, preventing unsafe behavior due to eagerness in speaker arbitration.
• Hybridization of modular and end-to-end strategies, leveraging FSMs for control while end-to-end models internalize turn-taking and temporal dynamics via large-scale synchronized data.

A projected path forward involves embracing next-token-pair prediction and dual-branch architectures to overcome the inherent sequential blocking of present autoregressive models in full-duplex speech-to-speech interaction systems (Ge et al., 26 Sep 2025). This will enable more humanlike, proactive, and seamless interactive systems for both wireless and dialogue domains.

7. Cross-Domain Extensions: Full-Duplex Beyond RF Communication

Full-duplex principles are generalizing into new domains including:

• Visible Light Communication (VLC): Devices, such as suspended InGaN/GaN MQW structures, simultaneously emit and detect light within a single chip, facilitating in-plane full-duplex transfer and paving the way for miniaturized, energy-efficient photonic circuits (Yang et al., 2016).
• Quantum Communication: Counterfactual protocols using electron-photon interaction gates demonstrate simultaneous classical or quantum information exchange without particle transmission, with the protocol behavior mapped to (quantum) erasure channels featuring capacity formulas

$$C = 2\zeta_c\ \mathrm{bits/Bell-pair} ,\quad Q = 2\max\{0,2\zeta_q-1\}$$

underpinned by the quantum Zeno effect for secure, full-duplex bidirectional communication (Zaman et al., 2019).

These advances exemplify full-duplex interaction as a unifying technical concept, informing both the most fundamental and the most applied realms of wireless communication, dialogue interaction, and emerging quantum and photonic systems.