Abstract
A spoken note can feel provisional while it is being said and acquire a different status the moment it becomes text. That shift is not captured by asking whether dictation is accurate, private, accessible, or productive in the abstract. It requires a closer question: when did speech become a draft, and what new audience, residue, or revision obligation arrived with it? This paper develops the construct of the speech-to-draft boundary: the deliberate handoff at which spoken material becomes a text artifact with different exposure, review, attribution, retention, and sharing conditions. Its instrument, the Dictation Boundary Protocol, records capture purpose, speaking environment, material class, capture route, transcription/handoff point, bystander or device exposure, transcript review, source/claim separation, retention or deletion condition, sharing threshold, accessibility question, and a mandatory non-conclusion. The Protocol is not a product score, security audit, legal opinion, medical recommendation, or rule against voice-based writing. It allows a writer, teacher, accessibility lead, or editorial coordinator to preserve the value of spoken capture while refusing to mistake a transcript for a reviewed, private, or shareable draft.
Scope
This is a teaching instrument for one situated modality handoff. It does not assess the security, privacy, accessibility, accuracy, or suitability of a named dictation system. It does not determine whether a workplace, home, vehicle, waiting room, or public setting is safe or unsafe for speech. It offers no medical or assistive-technology recommendation, no workplace/legal privacy advice, and no assurance that a transcript was deleted, retained, or never exposed.
The note that was supposed to disappear
The note begins as an experiment in not having to hold it.
In a clinic lobby, a researcher is waiting for a delayed appointment. A phrase returns: The archive did not simply preserve the record; it chose what could later look stable. She does not want to open the project file and make the idea official beside unchecked sources. So she speaks quietly into a phone, names a historian who may help, adds, “maybe that is the wrong source,” and says, “Do not turn absence in the record into evidence of intention.” The utterance feels like rehearsal. No paragraph, she thinks, has entered the manuscript.
That evening a transcript waits on the screen. The phrase, historian, and speculative clause have become text. They can be searched, copied, retained, exported, or forwarded. The researcher faces a question absent during rehearsal: When did this become a draft, and what new obligations arrived when it did?
The question is not whether she should have dictated. Speech recognition can be an essential route into text for people who cannot rely on keyboard or mouse input; it can also support dictation and interface control when another route is unavailable.[3] Nor is the question whether speech recognition is accurate “in general.” In a controlled smartphone study, Foley, Casiez, and Vogel treated composition and transcription as different tasks and separated preparation, input, and error measures.[1] Speaking a possible sentence, transcribing an utterance, and reviewing a draft are not the same work merely because each produces words.
The lobby does not settle the question either. A public room can contain strangers, staff, sound, and a writer’s uncertain sense of what is being overheard. But “public” is not a privacy diagnosis, and “private” is not an absence of residue. Research on social acceptability makes location and nearby audiences relevant variables for speech input without supplying a universal list of permissible settings.[4] Recording a condition is not a moral verdict. It is a reason to ask what is being spoken, who or what may receive it, and what it will become.
This paper calls the moment of conversion a speech-to-draft boundary: the deliberate handoff at which spoken material becomes a text artifact with different exposure, revision, attribution, and retention conditions. The boundary may occur when text appears live, when a transcript is saved or retrieved, when it is reviewed, or when someone else can receive it. It may be uncertain. A writer may not know the capture route, whether a history exists, whether a setting is active, or where text will remain. That uncertainty is not an invitation to invent a security claim. It is a condition the workflow needs to hold.
The shift matters because speech and text invite different actions. A writer may repeat a spoken phrase differently, abandon it, or let it dissolve. A transcript can be named, joined to a source, placed in a folder, copied into a manuscript, or left uninspected until its wording looks more settled than the thought ever was. Text has acquired persistence and a path toward editorial use. A protocol cannot erase that fact. It can make the handoff visible before a temporary rehearsal silently gains the status of evidence, prose, or a shareable note.
Speech is not yet a draft; text is not yet reviewed
It is tempting to describe dictation as a single operation: speak, then words appear. The description is useful for a keyboard shortcut; it is inadequate for editorial work. A single spoken capture event can contain several distinct crossings.
| Crossing | What changes | The editorial question |
|---|---|---|
| Purpose to utterance | A writer moves from intending to preserve, test, remember, command, or reflect to actually saying material aloud. | What kind of material is being spoken, and what has not yet been established? |
| Utterance to capture | Sound is taken up by a device, application, person, or unknown route. | What route is known, assumed, or unchecked? |
| Capture to transcript | Spoken language becomes visible or otherwise retrievable text. | Has a text artifact now been created, and where does it first appear? |
| Transcript to reviewed draft | A person inspects wording, separates a source from a speculative claim, and decides whether it says what it should say. | Has text been reviewed as text rather than merely received? |
| Reviewed draft to retained/shareable text | The material crosses a threshold into a note store, project file, collaborator channel, editor handoff, or publication system. | What new audience, residue, or editorial responsibility accompanies the next use? |
The table does not claim that all capture routes have the same technical behavior. It supplies a vocabulary for separating operations that common language collapses. The writer in the lobby has not merely “dictated a note.” She has formed a possible idea, spoken a named source, created or received a transcript, and postponed the review that would distinguish a prompt from an evidence-bearing sentence.
The distinction between composition and transcription is especially important. Foley and colleagues compared smartphone speech recognition and touchscreen typing in a controlled study that included both a constrained composition task and transcription. Their results concern particular short-phrase tasks and a particular studied configuration; they cannot tell us how a writer should develop a long manuscript. Their method nevertheless provides a useful discipline: composition has preparation and conception demands that are not identical to reproducing a given string.[1] A transcript may be mechanically legible while its thought remains uncomposed.
A second study helps keep capture situated rather than abstract. In a two-week deployment of a touch-and-speech personal-data-capture application, Luo and colleagues found that participants’ modality choices varied with data type, personal preference, error tolerance, and social surroundings; the authors reported different capture uses rather than a single preferred modality.[2] The finding does not make any individual account of dictation representative. It gives a reader reason to record the conditions around a capture event instead of assuming that voice is always either liberating or exposing.
The distinction is protective in both directions. It protects a writer from treating a provisional transcript as a reliable source note or polished claim. It also protects speech input from a familiar dismissal: that if it cannot complete an entire editorial chain without another mode, it has failed. A field note, a sensory observation, a sentence skeleton, a reminder to check a source, or a verbal outline can be valuable capture even if it requires a later review surface. The error is not changing modes. The error is failing to name when and why the mode change must occur.
The speech-to-draft boundary
A speech-to-draft boundary is not a line drawn by a device alone. It is a deliberate workflow handoff. The writer identifies the point at which an utterance has become a textual artifact that needs a new kind of care: visual or auditory review, source separation, a retention decision, an account of a possible audience, or a sharing threshold.
The boundary can be early. A writer may decide that the moment live text appears, it must be classified as an unreviewed prompt. It can be later. A writer may use speech to compose a phrase but not call it a draft until it has been moved into a manuscript and checked against its source. It can be interrupted. A writer may stop before transcription, decide that a named person or unverified claim should not continue through that route, and preserve only a neutral reminder. A boundary can also be unresolved: the capture route, history, or retention condition may be unknown. Unknown is a valid protocol value. It is more responsible than silently substituting a claim of local, private, deleted, or secure.
This construct does not turn a writer into a compliance officer. It gives the writer a moment of agency. Instead of asking, “Is dictation private?” the reader can ask narrower questions: What was I trying to capture? What kind of material did I speak? Where did it become text? What has been reviewed? What remains unknown? What must happen before this becomes a shareable editorial object?
The questions are deliberately conditional. A short reminder containing no source, name, or claim calls for a different boundary than a spoken passage that includes a person’s name, a tentative interpretation, a quotation fragment, and a next-step instruction. The protocol does not classify material as legally confidential or solve a security problem. It asks the writer to describe the material class in enough detail to select a proportionate next action.
The Dictation Boundary Protocol
The Dictation Boundary Protocol records one capture event and its first transition into a text artifact. It is most useful immediately after a situation that seems too small to merit a workflow: a phrase said while walking, a source reminder spoken between appointments, a voice-created email draft, a revision command, a reflection that arrived when a keyboard was inaccessible, or a transcript found after the fact.
| Field | What to record | Reader prompt | Boundary control |
|---|---|---|---|
| 1. Capture purpose | fleeting phrase, reflection, outline, source note, message, revision command, or another stated purpose | “What did I want this speech to do?” | Do not infer importance from length. |
| 2. Speaking environment | stated location, sound/social condition, and whether others may be co-located or unknown | “What environmental condition is relevant to this utterance?” | Do not label the setting safe or unsafe. |
| 3. Material class | public idea, named source, provisional claim, personal material, confidential material, or unknown | “What kind of material is being spoken?” | Do not make legal classifications. |
| 4. Capture route | named route, locally described route, online described route, unknown, or unchecked | “How did speech begin to become text?” | Do not infer the route from a device category alone. |
| 5. Transcription/handoff point | live text, saved transcript, queued output, manual transcriber, or unknown | “At what moment did a text artifact exist?” | Do not call every utterance a draft automatically. |
| 6. Bystander/device exposure | audible exposure, screen exposure, microphone/app permission, account/device history, or unknown | “Who or what might be implicated by this capture?” | Do not claim exposure without observation. |
| 7. Transcript-review step | visual review, auditory review, re-dictation, alternate correction, second surface, or deferred review | “How will wording be inspected before its next use?” | Do not mistake transcription for review. |
| 8. Source/claim separation | the boundary between a named source, a speculative thought, a quotation, and a fact claim | “What must be marked before it looks like evidence?” | Do not convert a spoken prompt into a source. |
| 9. Retention/deletion condition | stated history, export, deletion, unknown setting, retention period, or review needed | “What residue is known, unknown, or awaiting verification?” | Do not promise deletion or retention without verification. |
| 10. Sharing threshold | private note, collaborator, editor, publication system, or no sharing decision yet | “What must occur before another person receives this text?” | Do not treat saving as approval to share. |
| 11. Accessibility question and non-conclusion | access value at stake and what this record cannot establish | “What does this boundary preserve, and what remains unproven?” | Mandatory. The Protocol is incomplete without it. |
The Protocol does not demand that every field be fully known. A finished record may say: capture route: unchecked; retention condition: unknown; sharing threshold: no sharing decision; non-conclusion: this does not establish that any audio or transcript was retained beyond the visible note. Such a record is not weak because it contains uncertainty. It prevents uncertainty from being disguised as a comforting assumption.
The final field matters because a record can become retrospective blame. Lau, Zimmerman, and Schaub’s smart-speaker study identifies tensions among primary, secondary, and incidental users; not a dictation result, it keeps an audience beyond the speaker visible.[5] W3C identifies speech recognition as a means of dictating text and controlling an interface, noting it can be essential for some people.[3] The Protocol asks what access enables before setting a boundary. It can preserve capture while postponing review or sharing.
One note, three boundary readings
Return to the constructed lobby note. The same words change status at each crossing. The following field readings are illustrations, not reports about a real person, clinic, product, or disclosure incident.
| Protocol field | Reading A: spoken rehearsal | Reading B: transcript received | Reading C: editorial handoff later that night |
|---|---|---|---|
| Capture purpose | Preserve a phrase and a source lead before it disappears. | Recover what speech has already become. | Decide whether any wording belongs in the project record. |
| Material class | Provisional claim plus a named source and a self-correction. | Text containing a possible claim, a name, and uncertainty. | Candidate prompt, not evidence or publishable prose. |
| Handoff point | Not yet known to the writer; she experiences the event as rehearsal. | Text is visible and retrievable. | Text is moved only after review, or deliberately left outside the project. |
| Review step | None; the task is capture. | Mark the source lead and speculation as unreviewed. | Check the source, rewrite or remove the claim, and verify that the prompt is not cited as evidence. |
| Retention condition | Unknown or unchecked. | Record visible history/setting if known; otherwise retain “unknown.” | Decide whether the reviewed note stays, moves, or is deleted under an identified condition. |
| Sharing threshold | No sharing decision. | No sharing decision; a transcript is not collaborator-ready. | A shareable draft exists only after stated editorial review. |
| Non-conclusion | This does not show that anyone heard the utterance. | This does not show how the service retained or handled the transcript. | This does not prove the initial spoken claim was true or ready to circulate. |
Reading A is not an inferior kind of writing because it is unfinished. The purpose is preservation of a fleeting thought. Reading B is not automatically a problem because a transcript exists. It is a change of condition: material that felt rehearsed can now be copied and treated as text. Reading C does not require the writer to destroy the note or abandon dictation. It requires the writer to decide whether the note has crossed into editorial use and, if so, to review the relation between the named source and the speculative claim.
The difference between these readings is the heart of the action protocol. A workflow can designate an unreviewed prompt status. That modest label does several things at once. It preserves the access benefit of fast capture. It blocks a transcript from silently becoming a source note. It tells a future reader—often the same writer—that the wording contains a thought path, not a settled conclusion. It also creates an intelligible next step: review on a chosen surface, mark the claim/source relation, then decide whether the text should be retained, moved, shared, or discarded.
Review is a mode change, not a correction afterthought
A transcript is not reviewed merely because it is visible. The question is not whether every speech-recognition output will contain an error. The question is whether the writer has inspected the text for the purpose it will next serve. If a phrase will remain a private prompt, a quick marking step may be proportionate. If it will become a claim in a research note, a source citation, an editor handoff, or a sentence that affects another person, a different review burden may be appropriate.
Research on voice-based editing makes the distinction concrete. Ghosh and colleagues studied two eyes-free approaches to editing dictated text: command-like instructions and re-dictation over an erroneous portion. Their study found different strengths by the complexity of the edit and treated correction as a design problem after text formation, not as a negligible final click.[6] The finding does not decide how any writer should correct a transcript. It establishes that capture and revision should not be collapsed conceptually.
Sengupta and colleagues make a complementary point through a controlled comparison of voice-only and gaze-plus-voice approaches to selecting and correcting text. Their research system and desktop eye-tracking setup are not a smartphone recommendation. Yet the work shows why post-capture validation has its own interaction demands: locating a problematic word, selecting it, and choosing a correction are separate from producing an initial utterance.[7] The Dictation Boundary Protocol therefore permits a writer to choose second surface, alternate correction, or deferred review rather than demanding that all repair occur through speech at the moment of capture.
This is a useful relief from perfectionism. A writer does not have to create final prose in a waiting room. The responsible action may be to capture an idea, label it, and hand it to a later review mode. A teacher can make this visible in an assignment: spoken capture is allowed; any transcript inserted into a research draft must be labeled, checked against sources, and reviewed before it is submitted as argument. The instruction does not punish dictation. It distinguishes a productive first handoff from a completed editorial act.
Data route is part of the handoff, even when it is unknown
A protocol about speech-to-draft boundaries cannot assume that all speech capture shares one data route or one retention condition. It also cannot turn primary documentation into an advertising brochure or a security certification. The teaching use of documentation is narrower: it demonstrates why a writer should record the actual route, setting, history, endpoint, or unknown condition relevant to one capture event.
| Documentation record | Bounded instructional lesson | What it does not establish |
|---|---|---|
| Google Cloud Speech-to-Text documentation distinguishes named streaming/synchronous and asynchronous routes, describes an opt-in data-logging program, and states different transcript-handling conditions for those routes.[8] | “Speech-to-text” can hide different service paths; a protocol must name the relevant route or record it as unknown. | The behavior of a consumer keyboard, an integrated application, or every cloud service. |
| Apple’s Improve Siri and Dictation documentation describes an opt-in setting, request history, stated audio/transcript/request-data conditions, and user-facing controls for named features.[9] | A writer should not infer that a visible transcript is the only possible record; settings and histories can be part of a handoff inquiry. | A cross-platform policy, a security certification, or a reason to enable or disable a particular setting. |
| Microsoft’s Windows documentation distinguishes device-based and online recognition and describes voice typing, contribution settings, and permissions for named functions.[10] | Local/online route and optional contribution conditions are facts to record when known, not qualities to assume from the word “dictation.” | The processing route, retention behavior, or privacy of a different system. |
The table is intentionally not a buyer’s guide. It teaches a question with three honest possible answers: known, unknown, and not yet checked. A writer who does not know the route has not failed the Protocol. The writer has learned that a boundary is unresolved. The next action may be to check relevant documentation, use a neutral capture, defer named material, or leave the text outside a shareable project until a review step occurs. The Protocol cannot verify the route. It can prevent the writer from building a claim on an unverified route.
This distinction also prevents a familiar reversal: treating technical detail as the only form of privacy care. A fully documented route does not make a speculative claim reviewed, a named source accurate, or a draft ready to send. Conversely, a writer may use an accessible voice route responsibly while still recording uncertainty about a history or service condition. The speech-to-draft boundary contains both material residue and editorial obligation. Neither cancels the other.
Access is a reason to draw a boundary, not a reason to prohibit speech
A poorly designed privacy conversation can make voice-based drafting look like a careless alternative to “real” keyboard work. That is not the position of this paper. Speech recognition may be essential for some people and useful in circumstances where keyboard or mouse input is inaccessible, painful, impractical, or unavailable.[3] The appropriate response to that fact is not to withhold a boundary. It is to make the boundary adjustable.
A writer may use voice for a first phrase but choose a later review surface. A writer may decide that one material class—named sources, for instance—needs a different capture route, while ordinary reminders remain eligible for dictation. A writing program may permit spoken first drafts and require that any transcript used in an evidence-bearing assignment carry an unreviewed-prompt label until source/claim separation occurs. An editorial coordinator may state that spoken notes can enter a shared process only after a named review step. These are workflow possibilities, not prescriptions. Their value lies in allowing a boundary to preserve speech access rather than demand a universal form of silence.
The counterweight is real. The NoteWordy study did not find that speech and touch were interchangeable; participants used them differently across material and social conditions.[2] Pandey and colleagues’ research on audible and silent speech input likewise treats social acceptability as context-dependent rather than a fixed trait of a location or person.[4] A protocol should keep that variability visible. It should not ask a writer to justify using speech, and it should not pretend that a phone, a room, or a transcript has the same meaning in every composing practice.
Limits of the instrument
The Dictation Boundary Protocol does not measure speech-recognition accuracy, writer productivity, confidentiality, disability, expertise, or editorial quality. It cannot determine who heard an utterance, where a transcript resides, whether a system retained data, whether a setting met a legal obligation, or whether another person will be affected by a text. It does not certify accessibility, security, privacy, or compliance. It does not decide whether a writer should use voice, touch, a keyboard, an assistive technology, or another surface.
Its constructed lobby incident is an illustration, not a study. Its primary provider documents are named, time-bounded records, not general rules for every system. Its HCI studies use their stated participants, setups, task types, and research prototypes; they are evidence for distinctions and questions, not rankings of tools or predictions of individual experience.
The Protocol’s narrowness is a protection. A spoken note can become evidence of a writer’s supposed carelessness, a product’s supposed insecurity, a room’s supposed unsuitability, or a universal problem with voice input before anyone has described the material, route, audience, review step, or unknown. The record interrupts that speed. It lets a writer say: this was a source lead, it became visible text here, the review has not happened yet, the retention condition is unknown, and no sharing threshold has been met.
Conclusion: name the handoff before naming the failure
Speech can capture a thought when a keyboard cannot. A transcript can preserve a phrase that would otherwise vanish. Neither fact makes it a reviewed, evidence-bearing, private, or shareable manuscript fragment. Those statuses arrive through different handoffs.
The practical task is therefore not to declare dictation good or dangerous. Before assigning failure to a writer or service, record the capture purpose, material class, route, transcription point, review step, residue condition, and sharing threshold. State what remains unknown. The Dictation Boundary Protocol makes that modest discipline usable.
At what moment did speech become a draft, and what new audience, residue, or revision obligation arrived with it?
References
[1] Foley, Margaret, Géry Casiez, and Daniel Vogel. 2020. “Comparing Smartphone Speech Recognition and Touchscreen Typing for Composition and Transcription.” Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20). https://doi.org/10.1145/3313831.3376861
[2] Luo, Yuhan, Bongshin Lee, Young-Ho Kim, and Eun Kyoung Choe. 2022. “NoteWordy: Investigating Touch and Speech Input on Smartphones for Personal Data Capture.” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 6 (4). https://doi.org/10.1145/3567734
[3] W3C Web Accessibility Initiative. 2024. “Speech Recognition.” Web Accessibility Perspectives. https://www.w3.org/WAI/perspective-videos/voice/
[4] Pandey, Laxmi, Khalad Hasan, and Ahmed Sabbir Arif. 2021. “Acceptability of Speech and Silent Speech Input Methods in Private and Public.” Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI ’21). https://doi.org/10.1145/3411764.3445430
[5] Lau, Josephine, Benjamin Zimmerman, and Florian Schaub. 2018. “Alexa, Are You Listening? Privacy Perceptions, Concerns and Privacy-Seeking Behaviors with Smart Speakers.” Proceedings of the ACM on Human-Computer Interaction 2 (CSCW). https://doi.org/10.1145/3274371
[6] Ghosh, Debjyoti, Can Liu, Shengdong Zhao, and Kotaro Hara. 2020. “Commanding and Re-Dictation: Developing Eyes-Free Voice-Based Interaction for Editing Dictated Text.” ACM Transactions on Computer-Human Interaction 27 (4): Article 28. https://doi.org/10.1145/3390889
[7] Sengupta, Korok, Sabin Bhattarai, Sayan Sarcar, Scott MacKenzie, and Steffen Staab. 2020. “Leveraging Error Correction in Voice-Based Text Entry by Talk-and-Gaze.” Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (CHI ’20). https://doi.org/10.1145/3313831.3376579
[8] Google Cloud. 2026. “Data Usage FAQ — Cloud Speech-to-Text.” Updated 11 August 2026. https://docs.cloud.google.com/speech-to-text/docs/v1/data-usage-faq
[9] Apple. 2026. “Improve Siri and Dictation & Privacy.” Updated 11 February 2026. https://www.apple.com/legal/privacy/data/en/improve-siri-dictation/
[10] Microsoft Support. 2026. “Speech, Voice Activation, Inking, Typing, and Privacy.” https://support.microsoft.com/en-us/windows/privacy/speech-voice-activation-inking-typing-and-privacy