Interpretation Without a Second Device: The Single-Sided Assisted Interpretation Interface for Wearable Devices

Shijie Qu

KEOI Research Institute

sj@keoi.org

KEOI Working Paper · October 2026

Abstract

Assisted interpretation has an asymmetry problem. Existing devices assume either that a device is passed back and forth between the two parties, or that both parties wear one — split-pair translation earbuds, for example — and both impose coordination costs on the interlocutor, who in many real situations cannot or will not wear or operate anything. A second problem concerns delivery: when a device does suggest what to say, existing systems present text that the wearer must read silently and reproduce from memory, an error-prone route for unfamiliar phonetics. This paper describes an interaction interface, disclosed in U.S. Provisional Patent Application No. 64/170,248, in which only one party wears a device and the other does nothing at all. Candidate utterances in the interlocutor's language are shown numbered on a near-eye display, and each is also played once as in-ear text-to-speech audio through an open-ear speaker, so the wearer learns the pronunciation by shadowing — repeating what was heard — rather than by decoding romanized text. The wearer selects a candidate silently, by an electromyography (EMG) pinch gesture whose pinch count maps to the candidate's number, and reads the selected utterance aloud. If no selection is made within a configurable window, a fallback path lets the wearer speak natively once and read aloud the projected translation. Voice activity detection advances the interface state automatically, so the wearer never switches modes by hand. The paper reports no user studies and no measurements; the interface is described as disclosed, and its rationale presented as argument, not experimental result.

Keywords: single-sided interpretation; wearable interface; audio shadowing; electromyography gesture input; silent selection; cross-language conversation; human–computer interaction

1. Introduction

Interpretation assistance is usually designed as if both parties to a conversation were equally available to be equipped. In practice they are not. The person who wants assistance is motivated: they bought the device and put it on. The interlocutor — a business counterpart, a station attendant, a new acquaintance — has no such motivation, and in many encounters cannot be asked to acquire one. The disclosure states the two incumbent assumptions plainly: either interpretation is turn-based device sharing, a phone passed back and forth across the turn boundary, or it is symmetric deployment, both parties wearing earbuds, as with split-pair translation earbuds. Both impose coordination costs on the interlocutor — handling an unfamiliar device, surrendering an ear, synchronising with a machine they did not choose — and in a business meeting, a chance encounter, or a service interaction, the interlocutor may be unable or simply unwilling to pay them. A system that requires those costs fails at the moment it is most needed.

There is a second, quieter problem on the wearer's side. When an existing system does propose what to say, it presents text: the wearer must read it silently and reproduce it aloud from memory. For unfamiliar phonetics — a speaker whose native orthography gives no purchase on the target language's sounds — that route is, as the disclosure puts it, error-prone: the wearer has decoded symbols, not heard a pronunciation, and must guess at the prosody. What is missing is an interface that delivers the utterance to the ear, not only to the eye, and lets the wearer act on it without visibly consulting a machine.

This paper describes such an interface: the single-sided assisted interpretation interface disclosed in U.S. Provisional Patent Application No. 64/170,248 (filed October 6, 2026). Its constraint governs everything else: throughout the exchange, the interlocutor wears no device and performs no action. The wearer's device presents candidate utterances numbered on a near-eye display; plays each candidate once, in sequence, as in-ear text-to-speech (TTS), so pronunciation is acquired by shadowing; accepts a silent, non-verbal selection — in the preferred embodiment, a pinch gesture detected by an electromyography (EMG) wristband, with pinch count mapped to candidate number; confirms the selection; and leaves the wearer to read the chosen utterance aloud, as a participant in the conversation rather than the operator of a device. If the wearer selects nothing, a fallback path takes one natively spoken utterance, translates it, projects the rendering, and the wearer reads that aloud instead.

The paper's scope is deliberately narrow, matching the disclosure's: the interaction layer only — how candidates are presented, heard, chosen, confirmed, and spoken, and how the interface behaves when they are not chosen. How candidate utterances come into being is outside the disclosure this paper rests on, and nothing here depends on it. Sections 2–8 set out the model in sequence: requirements, flow, the shadowing rationale, the selection mapping, the fallback, state management, and alternative embodiments; Section 9 discusses limits, including the evidential status of the design rationale itself.

2. Design Requirements: The Asymmetry Constraint

Four requirements follow, each traceable to the situation described above.

R1 — The interlocutor does nothing. No device worn, nothing held, no action performed. Any element that leaks a demand across the table — a pause while the wearer fiddles, an earbud offered — violates the motivating constraint. The interaction must be completable entirely within the wearer's own sensory and motor space.

R2 — Pronunciation is delivered to the ear. Because the wearer's difficulty is phonetic as much as lexical, the candidate must be heard, not only seen. The delivery channel is in-ear audio at low volume, audible only to the wearer; and because the wearer must simultaneously remain a listener, the speaker is open-ear and non-occluding, preserving awareness of the interlocutor's speech and of ambient sound.

R3 — Selection is silent. Choosing what to say must not itself become a conversational act. A spoken command ("number two") spends the very channel the wearer is trying to perform in; a visible search through a menu advertises the machinery. Selection should be a small, private gesture, mapped onto the candidates by a scheme held in memory without looking it up — and confirmed within the same private channels, so the wearer knows the choice registered without announcing it.

R4 — Abstention must be graceful. Any interface that offers candidates will sometimes offer none the wearer wants, or the wearer may hesitate past the moment. Non-selection must be a normal branch, not a failure: there must remain a way for the wearer to say their own sentence — spoken natively, rendered by the system, read aloud — so the floor is never lost to the interface; and there must be an exit from the candidate machinery altogether when the wearer wants only to listen.

3. The Interaction Flow

The flow (FIG. 1 of the filing) begins at candidate availability — the moment a set of candidate utterances in the interlocutor's language is ready for presentation — and proceeds as a fixed, short sequence in which the wearer's part is mostly listening.

Presentation. The near-eye display shows the candidates — two to three in the disclosed embodiment — numbered 1 to 3. Numbering is not decoration: it is the address space the selection gesture will refer to, so sight and selection share one scheme from the start.

Audition. The open-ear speaker plays each candidate's TTS rendering once, in sequence, at low volume audible only to the wearer. The single pass is deliberate: the audition is a pronunciation model to be shadowed, not a loop to be studied, and its duration stays proportional to the utterance itself.

Selection and confirmation. The wearer selects silently by gesture (Section 5). The selection is confirmed by a brief display highlight and an optional confirmatory audio tick — feedback confined to the wearer's private channels, satisfying R3: the wearer knows the choice registered; nobody else does.

Utterance. The wearer then reads the selected candidate aloud to the interlocutor. Note what the interface has not done: it has not spoken. The voice the interlocutor hears is the wearer's own, producing an utterance the wearer chose and has just heard modelled; the conversation's surface is unbroken, and the interlocutor has been asked for nothing (R1). Throughout, the flow's tempo is set by speech rather than menus, and its attentional geometry is one-sided: display at the wearer's eye, audio at the wearer's ear, gesture at the wearer's wrist.

4. Audio Shadowing Instead of Romanized Text

The decision to deliver pronunciation as audio, rather than relying on the displayed text, is the interface's most deliberate design choice, and this paper presents it exactly as the disclosure does — as a reasoned design argument, not a measured result; no user study is reported in the disclosure, and none is claimed here.

A displayed candidate reaches the wearer as orthography. Where the target language's writing system is the wearer's own, orthography carries pronunciation with it, more or less. Where it is not — the disclosure's example is a Chinese speaker producing Japanese — the text must first be transcoded into a familiar notation (romaji, in that example) and then decoded into sound. Each transcode is a place for error, and the errors concentrate exactly where the wearer is weakest: in phonetic detail and prosody, which romanization represents coarsely or not at all, since written symbols underdetermine stress, timing, and intonation.

Hearing the utterance once, by contrast, delivers the phonetics and the prosody directly, as a model to be imitated. Shadowing — repeating after hearing — is the natural human mechanism for this: the wearer does not decode the candidate into sound but copies a sound just heard. The audition step (Section 3) is built around exactly this: candidates are played once, in sequence, so each is available to imitation in the moments before the wearer must speak. The display still carries the candidate's identity and number, but the burden of pronunciation is deliberately shifted from the eye to the ear.

The open-ear delivery completes the argument: an occluding earpiece would isolate the wearer from the conversation they are trying to join, whereas the disclosed speaker is non-occluding, so the shadowing model arrives inside the wearer's ongoing auditory scene rather than replacing it.

5. Silent Selection by Pinch Count

Selection is where an assistance interface most easily betrays itself: choosing is an act performed mid-conversation, and most input modalities make it visible or audible. The disclosed interface answers with a gesture chosen for privacy and a mapping chosen for memorability. In the preferred embodiment, selection is an EMG pinch gesture detected by a wristband. The mapping (FIG. 2 of the filing) is by pinch count to candidate index: a single pinch selects candidate 1, a double pinch candidate 2, a triple pinch candidate 3. Its virtue is isomorphism with the presentation: the candidates are numbered, the gesture counts, and the wearer who can count to three holds the entire scheme. There is no menu to navigate and no cursor to steer — the number of the candidate is the number of the gesture, and a count-based mapping that would collapse under ten candidates is well matched to two or three.

The gesture's social properties matter as much as its ergonomics: a pinch of the fingers is small, silent, and unremarkable; it uses neither the voice (reserved for the utterance itself) nor the gaze (which belongs to the interlocutor), and it can be performed while the wearer is still listening to the interlocutor's turn. Feedback closes privately: the highlight and optional tick (Section 3) tell the wearer, and only the wearer, that the selection registered.

One scope note belongs here: this paper describes the gesture at the level the disclosure specifies it — an EMG pinch detected by a wristband, mapped by count — and no deeper. The same channel also carries the interface's two marginal controls: a dismiss gesture that routes the flow to the fallback path (Section 6), and a long pinch that forces subtitle-only mode (Section 7); selection, abstention, and exit are all reachable without voice or menu.

6. The Fallback Path: Free Expression

An interface that only ever offers candidates would make the wearer a curator of other sentences — fluent exactly when the offered set happens to contain their thought, stranded when it does not. The fallback path (FIG. 1 of the filing) is the disclosed answer: it converts the interface from a suggestion box into an instrument the wearer can always speak through.

The path is entered on either of two triggers: the wearer makes no selection within a configurable time window, or issues a dismiss gesture, declining the set immediately rather than waiting out the window — whose length the disclosure leaves configurable, since the right patience differs between a rapid service exchange and a deliberative meeting.

In fallback, the wearer speaks natively, once. Speech recognition — on-device or in the cloud, as the disclosure provides — and machine translation produce the target-language rendering, which is projected onto the display; the wearer reads it aloud. The system then returns to listening for the interlocutor's next turn.

The fallback's design significance is disproportionate to its length in the flow. First, it guarantees expressive completeness: anything the wearer can say in their own language remains sayable; the wearer's repertoire is never reduced to the candidate set. Second, it disciplines the candidate mechanism: because abstention costs only one natively spoken sentence, there is no pressure to accept a poor candidate for fear of silence — candidate quality is gated by the wearer's own judgement, exercised freely. Third, the fallback divides labour between ear and eye asymmetrically: in the candidate path, pronunciation arrives by audio shadowing; in fallback, the rendering arrives as projected text. The disclosure does not extend the audition step to the fallback rendering — the fallback privileges immediacy over pronunciation support, so a wearer whose own sentence is the one they most need to say well gets the least phonetic help with it.

7. Turn-Taking and State Management

An interface with modes makes the wearer track the machine's state as well as the conversation. The disclosed interface instead drives its state from the conversation itself.

In the preferred embodiment, voice activity detection on the interlocutor's speech stream automatically advances the interface state through listening → candidates ready → awaiting selection → speaking. The states correspond to the turn's own phases: the interlocutor is speaking; candidates have been presented and auditioned; the wearer is deciding; the wearer is speaking. Because the advancing signal is the interlocutor's voice activity, the state machine's clock is the conversation's clock, and the wearer never manually switches modes — their attention is reserved for the two judgements only they can make: which candidate, if any, to say, and how to say it.

Automation of this kind needs a manual counterweight, and the disclosure provides one: a long pinch forces subtitle-only mode, imposed through the same silent channel used for selection. The override complements R4: fallback handles the turn where no candidate serves, and the override handles the stretch of conversation where none should be offered.

8. Alternative Embodiments

The disclosure records variations that mark the interface's intended range; each preserves the single-sided constraint while changing one channel of the flow.

Alternative selection channels. Selection may use temple touch, a voice command ("number two"), or eye-gaze dwell, in place of the EMG pinch. The pinch-count mapping is the preferred embodiment, not the definition of the interface: what the design requires is a selection mapped onto the numbered candidates, and the disclosure treats the channel as substitutable.

Display-free form. The wearable may be earbuds with no display, in which case candidates are presented purely via in-ear TTS with numbered audio prompts. The audition step then carries the whole presentation: the numbering the display made visible is spoken instead. That the flow survives removal of the display shows the design's centre of gravity — the auditory channel is primary; the near-eye display is its complement, not its foundation.

Voice-cloned TTS. The TTS voice may be cloned to match a preferred speaking style. The disclosure states this as a possibility and no more; this paper records it at that level. No implementation, and no voice data of any kind, is described in the disclosure or here.

9. Discussion and Limitations

This paper has described an interaction design, and its claims should be weighed as such. Three limitations bound them.

First, nothing here is measured. The paper reports no user studies, no selection or pronunciation accuracy, and no usability or accessibility evaluation; the disclosure contains none. The numerical values that appear — candidate sets of two to three; pinch counts of one to three; a selection window described only as configurable — are the disclosure's examples and design parameters, not experimental findings.

Second, the shadowing rationale is an argument, not a result. The disclosure's claim — that hearing prosody and phonetics once yields far more accurate reproduction than decoding romanized text — is plausible and motivates the whole arrangement, but within this body of work it is unevaluated, and this paper has phrased it as the disclosure's argument wherever it appears.

Third, the selection channel is specified only at the gesture level. The disclosure claims an EMG pinch detected by a wristband, mapped by count; it says nothing about recognition robustness, false activations, or performance across wearers, though the interface's viability plainly depends on such properties. They are questions the disclosure — and hence this paper — leaves open.

A final observation is interactional rather than technical: the interlocutor experiences only a person who listens and then speaks — the assistance is invisible by intent (R1). Whether such invisible assistance carries any obligation of disclosure, to a business counterpart say, is a norm deployment contexts will have to settle; the design settles only that the question will arise.

10. Conclusion

The single-sided assisted interpretation interface is an exercise in designing for an absent user: the interlocutor, who will wear nothing, hold nothing, and do nothing, and whose non-participation is the constraint from which every mechanism follows. Candidates are numbered so a gesture can address them; played aloud once so pronunciation can be imitated rather than decoded; selected by a count of pinches so the choosing stays silent; confirmed in private; and abandoned freely, through a fallback that returns the wearer's own voice whenever the offered words are wrong. State advances on the interlocutor's voice, and a long pinch hands control back at any time. What the design claims is modest and specific: interpretation assistance can be arranged so that its entire cost falls on the party who chose it, and its entire presence stays on one side of the conversation.

References

Levelt, W. J. M. (1989). Speaking: From intention to articulation. MIT Press.

Qu, S. (2026). Single-Sided Assisted Interpretation Interface for Wearable Devices. U.S. Provisional Patent Application No. 64/170,248, filed October 6, 2026.

Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696–735.

Starner, T. (2001). The challenges of wearable computing. IEEE Micro, 21(4).

Weiser, M. (1991). The computer for the 21st century. Scientific American, 265(3), 94–104.

AI Disclosure: This working paper was prepared with AI-assisted writing and editing tools. The design description, arguments, and conclusions are the author's own, and the technical content is drawn from the author's patent disclosure cited above.