Voice assistants have recently learned to talk and listen at the same time. They stop when interrupted, say "uh-huh" while the user is speaking, and answer within a few hundred milliseconds. Every one of these systems quietly assumes that the audio channel itself is full-duplex, meaning audio can flow in both directions at once. Many real links cannot do this. Push-to-talk radios, narrowband telemetry links, and many satellite and field radios are half-duplex: a transceiver cannot transmit and receive at the same moment. Put today's full-duplex agents on such a link and they go deaf whenever the far end is sending, which removes the barge-in and backchannel behaviour they were built for.
This note proposes to recover conversational full-duplexity on physically half-duplex links. The mechanism is time compression. Self-supervised speech tokens compress speech into discrete symbols at well under 2 kbps. At that rate, each side needs only a small fraction of the link's airtime to carry a second of speech. The link can therefore alternate direction in short, fixed cycles, and each receiver plays audio continuously from a small buffer. The research question is when this works and at what cost. Short cycles are limited by the radio's turnaround time. Long cycles add latency and break turn-taking. The tokenizer's bitrate decides how much slack the cycle has. We also propose to tell the dialogue model when the channel was blind instead of hiding it, and to test whether that restores turn-taking accuracy.
Core Claim
Consider a half-duplex link with raw rate $R$ (bits/s) and a fixed turnaround guard time $g$ (seconds lost each time the radio switches between sending and receiving). Speech is carried in both directions at bitrate $b$. In a repeating cycle of length $T$ with one slot per direction, each slot must carry $bT$ bits plus one guard interval. Feasibility therefore requires $2(bT/R + g) \le T$, which gives a minimum cycle length
$$T_{\min} = \frac{2g}{1 - 2b/R}.$$
The added one-way latency from buffering is on the order of $T_{\min}$. This bound follows from counting only airtime fractions. That step is justified because half-duplex throughput depends only on the fraction of time spent in each transmit/receive state, not on their order (Ong et al.). Ordering is therefore free for capacity, and the only cost of fine-grained alternation is the guard time. The expression blows up as $b \to R/2$. A conventional waveform codec whose bitrate is a sizable fraction of $R$ cannot emulate duplex at all, while a token stream with $b \ll R$ sits near the floor $2g$.
The claim has two parts. First, over the range of $R$ and $g$ where waveform codecs are infeasible or push latency past turn-taking tolerances, self-supervised speech tokens let an unmodified full-duplex dialogue agent keep most of its barge-in and backchannel accuracy. Second, the remaining loss is largely removed by feeding the link's blind intervals to the turn-taking controller as explicit input.
Existing full-duplex dialogue systems do not address this problem, and neither do semantic speech codecs. OmniFlatten interleaves speech and text tokens into one sequence to escape conversational "walkie-talkie" turn-taking (OmniFlatten). SALMONN-omni uses a streaming encoder and <shift>/<think> state tokens (SALMONN-omni). The neural finite-state-machine scheme emits speak/listen control tokens (A Full-duplex Speech Dialogue Scheme). All three assume a physically full-duplex audio path. On the codec side, LargeSC (Large Speech Model Enabled Semantic Communication) and Glaris (Error-Resilient Semantic Communication) carry discrete speech tokens over lossy packet networks, but they study one-way transport with random erasures, not two-way conversation under transceiver switching. The half-duplex relay coding work (Deep JSCC over the Relay Channel) handles physical half-duplexity, but only for static CIFAR-10 images. In this proposal, the tokenizer is the enabling mechanism, the half-duplex link is the constraint, and full-duplex turn-taking is the evaluation target.
Research Program
We will build a link emulator that imposes half-duplex operation with configurable raw rate $R$, guard time $g$, cycle length $T$, and residual packet loss. The emulator sits between a "user side" and an "agent side." On the user side, a small device encodes microphone audio into tokens, sends them in its slot, and decodes the agent's tokens to audio. On the agent side, tokens are decoded to waveform and fed to an existing full-duplex dialogue stack.
The tokenizers span a range of bitrates and designs. The first is HuBERT-style $k$-means units with repeated units removed, as in Direct Speech-to-Speech Translation With Discrete Units. The second is RepCodec, which learns to reconstruct self-supervised representations (RepCodec). The third is Mimi's residual quantizer truncated to its first few layers, as used in LargeSC. The fourth is the Glaris VQ-VAE with its hyperprior-based concealment. Each can be characterized by a measured $b$, and deduplicated units make $b$ variable. We will therefore also test a fixed-slot variant that pads to the worst case against one that lets slot length adapt.
We will then sweep the $(R, g)$ plane. At each point we will set $T$ to the smallest cycle that satisfies the bound with a safety margin, and measure three things: speech fidelity, dialogue timing, and turn-taking accuracy.
The second contribution is a schedule-aware controller. We will retrain a lightweight turn controller in the style of LPS-TC (Enabling Proactive Spoken Turns). It predicts normal turns, interruptive turns, backchannels, barge-ins, or no action every 40 ms. We will add two inputs. One is a per-frame channel-state token marking frames that arrived late or were buffered across a blind slot, similar in spirit to SALMONN-omni's explicit state tokens. The other is the known playout delay. Training data comes from WildTurn and synthetic dialogues, passed through the emulator at randomized $(R, g, T)$. The hypothesis is that a schedule-agnostic controller confuses jitter with pauses. It then takes turns early, or it responds to barge-ins a full cycle late. A controller that knows the schedule can discount this.
For single-relay links we will add one extension. Out-of-phase relays, which recover the half-duplex throughput loss (Fan et al.), keep one hop always active. We will ask whether this lowers the latency floor enough to matter for turn-taking at token bitrates.
Baselines
The do-nothing reference is push-to-talk: a half-duplex cascaded assistant that waits for end of speech, transcribes, generates, and synthesizes. It is the default on half-duplex radios today. A Full-duplex Speech Dialogue Scheme reports a first-token emission delay (FTED) of 2.28 s for its half-duplex baseline, against 0.68 s for its full-duplex model. We will run a push-to-talk version of the same dialogue model over the same emulated link. Any emulated-duplex configuration must beat it on FTED and on barge-in reaction time, with no loss in dialogue quality.
The strongest transport competitor is a conventional low-bitrate waveform codec run with the same time-division schedule. This is how time-division cellular voice already hides half-duplexity. It shares the same emulator, the same cycle-length optimizer, and the same dialogue stack; only the speech coder changes. We will use each codec's lowest intelligible setting, determined by measured WER, so that it is not handicapped. LargeSC reports that its token system reaches VisQOL 3.85 while operating at 550 bps to 2.06 kbps, far below the conventional codecs it compares against (LargeSC). We will test whether that gap translates into lower $T_{\min}$ and better turn-taking. Tokens beat this baseline if they give equal-or-better turn-taking F1 at a lower $R$, or lower barge-in latency at the same $R$.
The strongest dialogue reference is the unmodified full-duplex stack on an ideal full-duplex channel. This is an upper bound, not a competitor. On their own evaluations, SALMONN-omni reports F1 of 0.88–0.90 for separating real interruptions from noise and backchannels, and Freeze-Omni with LPS-TC reports an interruptive-turn F1 of 0.52 (SALMONN-omni, Enabling Proactive Spoken Turns). We will re-measure the chosen stack on our own test sets rather than compare against these numbers directly. Success means retaining most of the ideal-channel F1 at practical $(R, g)$.
The key ablation is the same token transport feeding the schedule-agnostic controller versus the schedule-aware one. The two use identical training data and compute and differ only in the channel-state inputs. This isolates the second contribution.
Why not just…?
Why not just send text? Running ASR on the user device and TTS at the far end would need even less bandwidth than tokens, and this is a real threat. The counter-evidence is indirect. The turn controllers that best separate real interruptions from backchannels and noise read audio, not transcripts (LPS-TC uses a Whisper encoder on raw dual-channel audio; SALMONN-omni uses a streaming audio encoder). The text-tape design is bounded by its 640 ms ASR chunking (A Full-duplex Speech Dialogue Scheme). We include text transport as a full baseline. If it matches tokens on backchannel and interruption F1, the honest conclusion is that tokens only buy voice fidelity and paralinguistics, and we will report it that way.
Why not just use a conventional codec with fast switching? Where $b \ll R$, this will work about as well as tokens, and the bound predicts it. The proposal's claim is limited to the region where the codec's $2b/R$ term pushes $T_{\min}$ past turn-taking tolerance. Glaris additionally reports that conventional codecs degrade badly under 30% packet loss (Glaris). The $(R, g)$ sweep will draw the boundary explicitly, and a result showing that boundary is narrow would be a useful negative finding.
Why not just keep push-to-talk and make the turn detector faster? Push-to-talk is blind to the other side for an entire utterance, not for one short cycle. Barge-in and backchannels are therefore impossible, whatever the detector's speed. The FTED gap of 2.28 s versus 0.68 s already measures how much latency the serialized design costs. The experiment still runs a push-to-talk arm with an aggressive end-of-turn detector so the comparison is not made against a straw man.
Proposed Experiments
We will evaluate turn-taking on WildTurn and duplex-dialogue-3k interactions replayed through the emulator. Metrics are FTED, interruption precision, interruptive-turn and backchannel F1, and barge-in reaction time. Reaction time is measured from the onset of the user's interruption to the agent's audio stopping at the user's ear. Transport fidelity is measured on LibriSpeech using WER and VisQOL, following the LargeSC and Glaris protocols. Dialogue quality is scored by an LLM judge, as in OmniFlatten.
The first decisive check tests the bound. Measured added latency should track $2g/(1-2b/R)$ across the sweep, and the codec baseline should hit the predicted wall as $2b/R \to 1$ while tokens do not. If measured latency departs strongly from the bound, the airtime model is wrong. That would most likely point to tokenizer lookahead or decoder chunking, which Textless Streaming S2ST shows trades against quality at small buffers. We would then add that term to the model.
The second check compares tokens, text, and codec at matched $R$ and $g$. It determines whether acoustic tokens justify their cost over text.
The third check is the controller ablation. If schedule awareness does not close a measurable part of the gap between emulated and ideal duplex on barge-in reaction time and interruptive-turn F1, that contribution is dropped. The result then reduces to a transport study.
References
- On Capacity and Optimal Scheduling for the Half-Duplex Multiple-Relay Channel
- OmniFlatten: An End-to-end GPT Model for Seamless Voice Conversation
- SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation
- A Full-duplex Speech Dialogue Scheme Based On Large Language Model
- Large Speech Model Enabled Semantic Communication
- Error-Resilient Semantic Communication for Speech Transmission over Packet-Loss Networks
- Deep Joint Source-Channel Coding Over the Relay Channel
- Direct Speech-to-Speech Translation With Discrete Units
- RepCodec: A Speech Representation Codec for Speech Tokenization
- Enabling Proactive Spoken Turns via a Generalized Style-Aware Full-Duplex Framework
- Recovering Multiplexing Loss Through Successive Relaying Using Repetition Coding
- Textless Streaming Speech-to-Speech Translation using Semantic Speech Tokens