Introduction
The person at the microphone can quote someone absent
A presenter says that Lina wrote about reopening a workshop, then reads the sentence aloud. The words belong to Lina within the presenter's account, but the voice heard in that moment may still be the presenter's. Replacing the turn label with Lina collapses the recorded speaker and the person being quoted into one role. That can make an absent person appear to have spoken at the event. The distinction is easy to see in a short illustrative sentence and easier to lose in a long transcript with names, quotation marks and several reported conversations. A label placed above the words carries its own claim about the recording. It should not be changed merely because the next sentence attributes a statement to another person. First establish who is producing the sound, then retain the reporting language that explains whose words or viewpoint the speaker is representing.
Start with the actual speaker transition. A moderator handing over to a presenter is one kind of change. The presenter switching from their own explanation to a quotation is another. The latter may change delivery, vocabulary or point of view while leaving the recorded participant unchanged. A real transcript needs evidence from the recording and any reliable identification available for the event. When that evidence does not support a name, a consistent provisional label is more honest than borrowing a name from the quoted material. The aim is not to remove quoted attribution. It is to preserve both relationships so the reader can tell who spoke at the event and who was represented in the speech. Those two questions are different even when the transcript uses the same person's name in both roles at other points.
References: TEI: Utterance; W3C WAI: Transcribing Audio to Text.
An authored speech makes the two layers visible
The control uses twelve fictional turns rather than a real event recording. Mara is assigned nine turns as the presenter, and Owen three as the moderator. Three other named people, Lina, Tomas and Nadia, appear only as absent people being quoted. There are six quotation records, including Mara quoting a previous answer of her own and a passage in which Lina is reported as quoting Tomas. The speaker assignments come from the authored scenario; the script does not discover them by analyzing sound. It checks textual relationships: each identifier resolves, each quoted span matches the indicated turn and the inner quotation falls within its recorded parent span. That limited experiment supplies a clear account of what is being preserved without pretending that a text-only check can establish the identity of a real voice or the historical truth of the statements represented.
The comparison deliberately tries a wrong reassignment rule. It replaces a turn's recorded-speaker field with the last quoted attribution target found in that turn. Four turn labels change, including the nested quotation. Counting all participant and attribution targets together yields five people rather than the two stipulated recorded speakers. Those are properties of this constructed example, not an observed model error rate. They demonstrate the consequence of losing a distinction in the data. The practical sections that follow use the example to examine speaker mapping, quotation boundaries, uncertain identities and handoffs to an editor or another tool. The central test is whether the result still describes the event the recording contains, while preserving what its actual speakers said about other people. A transcript should not create an additional event participant merely to represent a change in whose words are being discussed.
References: TEI: Speech or Thought; TEI: Attribution to a Person or Group.
1. Separate the Relationships a Label Can Hide
1.1 The voice, the name and the quotation have different evidence
A recorded participant, a speaker label, a personal name, a quotation, a reporting frame, a turn boundary and an identity claim are related without being equivalent. The participant produces speech in the recorded event. A label associates passages with that participant under the transcript's conventions. A personal name adds an identification that needs support. A quotation represents words attributed within the speech, while its reporting frame explains that relationship. A turn boundary marks how the transcription segments the event; it need not correspond to every change of topic or quoted perspective. In the authored example, Mara remains the recorded speaker through several quotations. Lina's name belongs in the report about her words, but its appearance does not make Lina the person speaking into the microphone. Keeping these roles separate prevents a formatting decision from adding an unsupported fact about participation.
For a real recording, establish a repeatable labeling basis. Listen across a suspected transition and consider reliable event identification where available. A program can help identify an introduced presenter, but a printed name alone does not establish who produced every later sound. Likewise, a change in sentence style or the appearance of another person's name is not sufficient evidence of a new recorded speaker. A transcript can retain provisional labels while the editor resolves a name. What matters is that the same evidence supports the label wherever it is used. If the recording or accompanying material does not resolve the issue, preserve the uncertainty. That is a more useful record than a confident name borrowed from a sentence whose purpose was to quote an absent person. It also leaves a clear question for later review rather than embedding the guess throughout the transcript.
References: TEI: Utterance; TEI: Attribution to a Person or Group.
1.2 A quotation can sit inside one participant's turn
The TEI speech guidelines distinguish an utterance and its participant attribution from other information represented within the text. Their examples include a reading shift within the same attributed utterance. TEI also provides a separate element for represented speech or thought, with context determining the relevant attribution. These encoding concepts help explain why a public speech can contain several represented voices without several new event participants. They are not an algorithm for recognizing those participants. The owned example uses ordinary JSON fields rather than claiming TEI schema conformance, but it keeps the same useful layer distinction: one field records the stipulated event speaker, and another records the person to whom a quoted passage is attributed. The two can agree, as in a self-quotation, or differ, as when Mara reads Lina's reported statement.
The practical editing consequence is to keep the reporting language visible. If a presenter introduces a quotation with a name and a verb such as wrote or asked, deleting that frame can make the following first-person words appear to be the presenter's own current statement. Adding a new speaker label creates a different error by suggesting an actual change of participant. Preserve the words and use clear quotation or editorial conventions appropriate to the transcript. A reader should be able to follow the shift in represented perspective without being misled about the sound source. This does not require an elaborate data format for every ordinary speech. It requires a consistent decision about what the label means and enough retained context to keep that meaning intact when the text is read, searched, summarized or handed to another editor.
References: TEI: Transcriptions of Speech; TEI: Speech or Thought.
1.3 Quoted first-person words do not automatically describe the presenter
A quotation's first-person pronoun belongs to its represented perspective, while the current turn label identifies the recorded event speaker. A reporting verb often carries the connection between them. In the fictional speech, Mara reports Nadia asking whether the workshop was open every day, then quotes her own earlier answer about Saturdays. The label remains Mara in both turns, but the quoted attribution differs. Treating every I as the current presenter's fresh assertion would lose the distinction between present explanation and reported past words. Treating every quoted attribution as a new event speaker would lose it in the opposite direction. The useful transcript retains the spoken frame and the relationship it establishes, then leaves any separate question about the truth of the reported account to evidence outside the act of transcription.
This matters when extracting statements from a speech. A passage under Mara's label can contain a claim she attributes to Tomas and then discusses or challenges. A later note should preserve who attributes what to whom, rather than present the quoted claim as Mara's unqualified position or as Tomas's direct participation at the event. The distinction remains necessary even if the wording sounds polished and the personal name is easy to recognize. Identification, attribution and endorsement are separate judgments. The authored example can demonstrate their representation because its reporting frames were deliberately supplied. A real editor should check those frames against the recording, including any uncertainty about whether the speaker said a name or pronoun. Do not repair an ambiguous attribution by silently choosing the person whose viewpoint seems most plausible from the subject matter.
References: W3C WAI: Transcribing Audio to Text; TEI: Transcriptions of Speech.
2. Follow the Authored Quotation Control
2.1 Two recorded roles remain separate from five named people
The fictional scenario defines Mara's nine turns and Owen's three turns before any quotation analysis. Mara is the presenter, Owen the moderator. The five-person register also contains Lina, Tomas and Nadia, who are not assigned recorded turns. Six passages have quoted attribution, and four distinct people are attribution targets because the set includes Mara's self-quotation. These counts answer different questions. Two is the number of stipulated event speakers; five is the number of registered people; four is the number represented as sources of quoted words; six is the number of quotation records. The separate fields make the differences explicit. Nothing in the mechanical check turns a text name into an acoustically identified person. It verifies the relationships supplied by the author and exposes what happens when a deliberately wrong transformation discards them. The unquoted mention in T09 provides another boundary. Mara refers to Lina's account there, but the annotation creates neither a quotation record nor a new speaker turn for Lina. A person reference can identify the topic of a sentence without claiming that the following words reproduce that person's speech. The register therefore should not be read as a list of voices detected in the recording, and the quote table should not be read as a list of every name mentioned. Each collection has a defined role, which determines what its count means.
WhisperAITranscribe, an online audio and video transcription tool, can provide a text stage for a recording workflow. This control did not test its speaker-label accuracy or a native quotation-aware feature. Its relevance is the review applied to any output: keep the label associated with the person producing the recorded speech, and preserve the quoted attribution within that speech. Check the actual recording before accepting a change in participant or a personal name. If the generated text supplies a label without enough evidence to identify it, review that label rather than treating a nearby quoted name as the missing answer. The product category explains where text preparation fits; it does not establish the result of an unperformed speaker-recognition test. The example's two speaker roles remain stipulated properties of its authored script, not a performance claim about an audio tool.
References: TEI: Utterance; TEI: Speech or Thought.
2.2 A self-quotation tests the roles without adding a person
Mara's reported answer, only on Saturdays, supplies the self-quotation case. Its recorded speaker and attribution target both identify Mara, while the reporting time differs from the current act of speaking. A data model that stores only a personal name could miss this difference because the same identifier appears on both sides. The quotation relationship still matters: Mara is representing an earlier answer, not necessarily announcing a new schedule for the present event. The authored text makes the distinction explicit through its reporting frame. A real transcript should preserve whatever corresponding words are actually heard. An editor should not remove the frame on the assumption that it is redundant simply because the quoted and current speaker names match. The relationship carries information beyond the identity of the person.
This is also why speaker labels alone cannot answer every question about a public speech. A search for Mara's turns can correctly retrieve the self-quotation, a quotation of Lina and Mara's own explanation. The shared label identifies the event speaker, while the passages have different attribution structures. A summary or evidence note needs to examine those structures before turning them into statements about what Mara currently believes, promises or reports. The control does not decide the historical truth of her earlier answer. It shows how to retain the fact that the current text presents it as an earlier answer. That is a smaller and more defensible claim. Keeping the distinction visible lets a later reader investigate the reported account without confusing the person who delivered the speech with the time and context of every statement represented inside it.
References: TEI: Speech or Thought; TEI: Transcriptions of Speech.
2.3 A nested quotation preserves a chain rather than a new turn
The nested passage has Mara producing the turn, Lina as the outer attribution, Tomas as the inner attribution and two linked quotation spans. Mara reports Lina telling her that Tomas said to keep the side door closed. The inner text lies within the outer quoted span, and both lie within Mara's authored turn. The mechanical check verifies that containment and resolves the referenced people. It does not confirm that Lina or Tomas actually made the statements in any real event. The deliberately naive rule chooses the last attribution target and replaces Mara's turn label with Tomas. That is one of four changed turn labels. It illustrates a concrete failure: flattening the quotation chain can turn the endpoint of a report into the apparent producer of the recorded passage, even though the scenario assigned the entire turn to Mara.
Preserve the chain when the distinction affects meaning. A note can say that the presenter reports Lina quoting Tomas, rather than reducing the passage to a direct statement by Tomas at the event. That wording identifies the evidence available in the recording without claiming access to an earlier conversation. It also preserves distance between the report and the underlying claim. A reader evaluating the statement can see which part was directly heard and which part was represented by that speaker. The owned case uses text spans to make the structure checkable, but those offsets are positions in a string, not timestamps in audio. A real workflow may use time references to help an editor return to the passage; it still needs a separate account of the attribution chain rather than assume that a timestamp or a name resolves the semantic relationship by itself.
References: TEI: Attribution to a Person or Group; TEI: Transcriptions of Speech.
3. Review the Recording Before Accepting a Label
3.1 Check the handover rather than the next proper name
Begin with the audible transition, the available identification evidence and the label's intended meaning. A moderator introducing a presenter may supply a useful starting point, but continue checking when the event changes speaker later. Listen before and after the suspected boundary, retaining enough surrounding context to hear whether a handover occurred or the same person began reading. A quoted name is a reason to examine attribution, not an automatic instruction to create a new participant. If a transcript already separates a quotation into a new labeled block, verify that block against the recording. The apparent neatness of the text cannot settle whether its segmentation describes the sound. An editor should be able to explain why the label changed at that point using the source, rather than point to the name that happened to appear in the sentence. Check the passage on both sides of an edit as well. If an excerpt starts halfway through a quotation, its opening words may lack the frame that made the original attribution clear. Returning to the longer source can resolve that missing context without inventing another voice. Where the longer source is unavailable, keep the limitation attached to the excerpt instead of treating the absence of an introduction as evidence of a new participant.
Keep the method proportionate to the recording. A straightforward introduced speech may have a small number of clear transitions. A panel, an audience exchange or an inserted recording may require more. Do not force a predetermined speaker count onto an event merely because the main presenter does most of the talking. Equally, do not increase the count for each person discussed. The authored example is intentionally simple so its relationships can be checked; it is not a template saying that every public speech has two participants. For a real recording, record the unresolved boundary or name and return to it with the relevant evidence. A stable provisional label can preserve continuity while that work is done. A guessed personal name can spread the original uncertainty into every later quotation, summary and search result that relies on the transcript.
References: W3C WAI: Transcribing Audio to Text; TEI: Transcriptions of Speech.
3.2 Preserve the words that begin and end the quotation
A reporting frame and a return to current speech help the reader understand a quoted passage. In the control, wrote, asked and my answer was distinguish different relationships without changing the recorded speaker. A real speaker may signal those boundaries less neatly, and punctuation in an automatic transcript may not reflect them correctly. Review the wording and the surrounding audio before deciding where the represented passage begins or ends. Do not add an attribution just because it would make the sentence easier to understand. If the source does not support a clean boundary, use an explicit editorial uncertainty appropriate to the transcript rather than manufacture a precise quotation. The aim is to preserve the relationship expressed in the recording, not to make every passage conform to the cleanest possible written dialogue.
Check what happens when the passage is excerpted. The words only on Saturdays can look like a current answer if separated from Mara's frame about an earlier answer. The inner door instruction can look like a directly heard statement by Tomas if separated from the report through Lina. Those are two distinct losses: time context and reporting distance. A useful extract retains enough of the frame to avoid them, or describes the attribution in an accompanying note clearly separated from the transcript. WAI's guidance on accurate transcription supports keeping the source's meaning intact; it does not make an editor's explanatory addition part of the recorded words. Maintain that boundary when improving readability. The text can be easier to navigate without silently changing who spoke, when the represented statement belongs or how the current speaker says they know it.
References: TEI: Attribution to a Person or Group; W3C WAI: Transcribing Audio to Text.
3.3 Keep uncertain speech and uncertain identity separate
An unclear word, an unidentified participant and an uncertain quotation target are different problems. A passage can be intelligible while the person's name is unknown. A known presenter can utter an unclear name inside a quotation. An audience response can be audible without supporting a confident individual identity. TEI's speech guidance allows representation of uncertain or grouped attribution, and WAI's transcription guidance addresses unclear words. These are ways to preserve limits in the record, not permission to fill them with plausible names. A transcript that says a participant is unidentified can still preserve useful words. A transcript that confidently labels that participant with a person merely mentioned nearby makes a stronger claim than the evidence supports and may mislead every later use of the text.
When reviewing text prepared with WhisperAITranscribe, resolve each kind of uncertainty on its own evidence. Check the audio for the wording and the transition; use reliable event identification where it applies; preserve an unresolved name or attribution when it does not. The product's output is a starting point for that review, not proof that an automatically supplied name has been independently verified. Do not merge two provisional labels merely because their quoted material concerns the same person, or split one label because a speaker adopts another perspective while reading. Those changes should follow the recording. If the task requires certainty that the available source cannot provide, state that limitation in the review record. A clear unresolved point is easier to investigate than a finished-looking label whose original basis has disappeared.
References: TEI: Transcriptions of Speech; TEI: Utterance.
4. Check the Annotation Without Overclaiming What It Proves
4.1 Structural checks can catch a broken relationship
The authored case checks identifier resolution, quoted text spans, parent containment and recorded-speaker assignments against its supplied data. Each of the twelve turn identifiers is unique. Every speaker and attribution target resolves to the five-person register. Each quoted span reproduces the indicated words in its source turn. The inner quotation's character range falls within the outer range in the same turn. These checks can catch a missing reference or a span attached to the wrong text. They do not establish the correctness of a real-world identity or the truth of the quoted claim. A perfectly consistent annotation could still encode a mistaken interpretation if its initial assignments were wrong. The value of the mechanical check is therefore specific: it preserves the relationships the editor has chosen and makes certain kinds of accidental damage visible during a handoff.
Keep the semantic review separate. Ask whether the source actually supports the speaker assignment, quotation boundary and reporting chain. In the control, the author supplied those conditions, so the script could compare against a known constructed answer. In a real event, the editor must obtain evidence from the recording and appropriate supporting material. A passing reference check cannot replace listening to a disputed transition. Nor can a name in a person register turn an uncertain voice into a verified identity. This separation prevents a useful technical result from being oversold. It also makes corrections more focused: a broken pointer calls for repairing the data relationship, while a mistaken attribution calls for revisiting the interpretation and any downstream passages that inherited it. Both can produce wrong output, but they require different investigative work. An additional authored counterexample changed the first quotation's target from Lina to Nadia while leaving its text span and turn reference untouched. All identifier, span and parent-containment checks still passed: Nadia existed in the register, and the quoted words still matched their location. The interpretation was nevertheless wrong for the supplied frame saying that Lina wrote them. That judgment came from reading the authored frame, not from a semantic model hidden in the check. This gives a concrete reason to review both kinds of evidence. Structural consistency can preserve an interpretation accurately while that interpretation itself needs correction.
References: TEI: Utterance; TEI: Attribution to a Person or Group.
4.2 Correct a label without transferring the quoted claim
For a transcript prepared with WhisperAITranscribe, keep a correction connected to the affected passage, the previous label, the revised label, the supporting evidence, the quotation layer and the remaining uncertainty. Suppose a later review establishes that an initially unidentified event participant is the moderator. Updating that participant's name should not automatically replace every attribution target inside their speech. The quoted people remain separate relationships. Conversely, correcting the name of a person quoted in a passage does not by itself change the recorded speaker. The owned example makes this easy to see because the speaker and quote fields can be examined independently. A real text editor may have only visible labels and punctuation, but the same reasoning applies: identify which claim is being corrected before applying a broad replacement that could alter the other layer.
Review the correction's reach. If a provisional label appears across several passages, confirm that they refer to the same recorded participant before changing all of them. If a personal name appears both as a speaker and as a quoted source, a text-wide find-and-replace can affect more than the intended role. Preserve the earlier reviewed version or a suitable change record so the basis of the revision remains inspectable. These are proposed editorial practices, not native capabilities tested in a transcription product. Their purpose is to prevent a small, justified correction from silently rewriting unrelated attribution. A useful change record explains what evidence improved and what relationship changed. It does not merely announce that the names were cleaned up, leaving a later reader unable to tell whether the edit concerned spelling, speaker continuity, quoted attribution or a substantive reinterpretation of the event.
4.3 Check what a summary says about participation
When using a WhisperAITranscribe transcript downstream, distinguish speech delivered at the event, words attributed within it, the current speaker's position and external verification. A summary saying Tomas spoke at the event would not fit the authored case, where Tomas has no recorded turn. A summary saying Mara reports Lina quoting Tomas would preserve the represented chain. A summary treating every quoted claim as Mara's own endorsement would introduce another unsupported interpretation. These differences can survive even when all personal names are spelled correctly. The review needs to examine the relationship carried by the sentence, not only its vocabulary. A statement can contain the right names and still assign the wrong role to each of them, which is why an accurate name list is insufficient evidence for an accurate account of the speech.
Use source-linked notes to keep those distinctions visible. A later question such as what Tomas said should be answered with the available evidence: the recording contains a report by its actual speaker, and the underlying statement may need separate verification if the task calls for it. Do not turn a quote embedded in a public speech into direct access to the earlier conversation. The same caution applies to an event attendance list inferred from a transcript. A person mentioned or quoted is not necessarily present, and a present person may have no spoken turn. The authored register deliberately separates these roles, but a real event may require additional evidence for attendance. Keep the answer tied to the question being asked. Speaker identification describes a relationship in the recording; it cannot silently certify every biographical or historical claim the speech makes about the people it discusses.
References: TEI: Speech or Thought; W3C WAI: Transcribing Audio to Text.
5. Preserve the Layers Through Later Work
5.1 Keep a speaker map and an attribution map linked but distinct
A useful future handoff could retain a recorded-speaker map, a quoted-attribution map and links back to the source passage. The two maps would answer different questions without forcing an editor to choose one at the expense of the other. In the control, the first identifies Mara and Owen across twelve turns. The second connects six spans with the people represented as their sources, including Mara herself and the two levels of the nested quotation. An application or working document could present those relationships more simply than the internal records, but it should preserve their meaning through export or editing. This is a proposed design principle, not a claim about a tested product format. Its practical acceptance test is whether the recipient can still distinguish the event participant from the person whose words that participant reports.
The handoff should also make uncertainty inspectable. A confident speaker association and an uncertain quoted name need not share one confidence judgment. An unidentified audience voice and a clearly reported absent person's name are different kinds of evidence. Keeping those distinctions separate can prevent a downstream tool from treating every populated name field as equally verified. The owned example contains stipulated roles, so it does not supply measured confidence values or a procedure for calibrating them. It supplies a reason to preserve the basis of each annotation. An editor receiving the work should know which decisions came from the recording, which came from reliable event material and which remain provisional. That is more useful than a single finished flag on a document whose speaker labels and quoted attributions were established by different methods.
References: TEI: Transcriptions of Speech; TEI: Attribution to a Person or Group.
5.2 Test an edit against the relationship it could damage
A focused maintenance check can name an editorial operation and a relationship to preserve. Renaming a provisional recorded speaker should leave quotation targets unchanged unless the evidence specifically changes those targets too. Extracting a quoted passage should retain its reporting chain where that chain affects interpretation. Moving text into a summary should not turn an absent quoted person into an event participant. The authored example can support small regression checks for these transformations because its expected roles and spans are known. Such checks would be controls for a chosen editing process, not proof that an audio tool recognizes speakers correctly. A real workflow still needs source review when the interpretation changes. The value of the regression case is to catch accidental loss of a distinction the editor has already decided to retain.
If WhisperAITranscribe supplies the initial text, apply that check to the actual handoff you use. Confirm what labels, quotation marks, source references or notes the retained artifact preserves, and review any transformation that flattens them. Do not infer a native structured export from the general category of transcription software. A simple text output can still be useful when the reporting frames remain clear and the editor keeps an adequate speaker map. A richer representation can still be misleading if it encodes an unsupported identity. Choose the representation for the task, then verify its content after the operation. The controlling question is whether the next reader can recover the intended relationships from the artifact they actually receive, rather than whether the originating interface once displayed a correct label before the text was copied, edited or summarized.
References: TEI: Utterance; W3C WAI: Transcribing Audio to Text.
5.3 Evaluate speaker features through the public speech task
A comparison involving WhisperAITranscribe should separate speaker continuity, personal-name identification, quoted attribution and editorial usability. WhisperAITranscribe can be considered for preparing the text, while any current speaker feature and its behavior need their own verification. A system may group recurring speech under labels without independently establishing the person's name. A transcript may preserve names in quoted material without correctly labeling the event participants. A polished summary may hide the difference by converting everything into short declarative sentences. Evaluate the result against the public speech task rather than assume that one attractive output proves all four relationships. The authored control offers no recognition benchmark, but it gives a concrete challenge an editor can look for: several represented people can appear inside one presenter's continuing turn.
For WhisperAITranscribe or any alternative, use a permitted recording representative of the actual work when evaluating a speaker feature, then inspect the relevant transitions and quoted passages. Keep the observed result separate from what the feature description led you to expect. If the task requires verified names, determine what evidence establishes them; do not credit anonymous clustering with a stronger identity claim. If the task is an accessible transcript for readers, preserve the distinctions needed to understand who is speaking and who is being represented. If later research or reporting depends on a quote, retain the source passage and the reporting chain. The useful outcome is a transcript that remains intelligible through those uses. A speaker label earns its value by preserving the event's relationships, not by making every proper name look like a participant.
6. Conclusion
6.1 Preserve the event and the represented account
A transcript prepared with WhisperAITranscribe should distinguish the recorded speaker, the personal-name evidence, the quoted source, the reporting frame, the quotation boundary and the remaining uncertainty. The authored control shows why those distinctions matter. Its twelve turns belong to two stipulated participants, while six quoted passages represent four attribution targets within a five-person register. Mara's self-quotation keeps the same person in two roles without making the earlier answer a new current statement. The nested quotation preserves a report through Lina to Tomas while Mara remains the event speaker. Flattening those relationships produces a different account of who participated. A correct speaker label cannot be derived simply by counting names, taking the last quoted person or treating a change in represented perspective as a handover at the microphone.
For a real recording, check the transition and the identification evidence, then keep quoted attribution inside the speech where it belongs. Preserve the words that introduce and qualify the quoted passage. If a name or boundary is uncertain, leave an explicit unresolved point rather than tidy it into a confident label. The text-only control can demonstrate consistency and the consequences of a wrong transformation because its intended roles were supplied. It cannot establish how a real recognizer would perform or whether a reported conversation happened as described. Keep those limits attached to the evidence. That leaves the editor with a practical method: establish who produced the recorded words, retain the relationships those words express and verify that later editing has not silently changed either layer.
6.2 Use four steps for speaker and quotation review
Step 1: Establish what the speaker labels mean for the recording and create a stable provisional map. Use the actual transitions and reliable identification evidence where available, leaving unsupported names unresolved. Step 2: If WhisperAITranscribe is used for text preparation, review its actual output against the recording. Check a suspected handover before accepting a new label, especially where the speaker begins quoting someone else. Step 3: Preserve the reporting frame, quotation boundaries and any nested attribution needed to understand the passage. Keep the event speaker separate from the person represented, including cases where both roles identify the same person. Step 4: Check the retained artifact and any summary or extract for the same relationships. A copied passage should not turn an absent person into a recorded participant or an attributed claim into the presenter's endorsement. The working goal is a traceable account of who spoke and who was represented, supported by the evidence actually available.
When a correction is needed, identify the relationship that changes. Updating a recorded speaker's name is different from correcting a name inside a quotation. Revising a quotation boundary is different from finding a new event participant. Keep a suitable record of the affected passage and the evidence so the edit can be understood later. A mechanical reference check can help preserve an accepted structure, but it cannot decide that interpretation on its own. If the source remains unclear, keep the uncertainty visible through the handoff instead of removing it for a smoother final page. This gives the next editor a defined problem to investigate and prevents a convenient formatting decision from becoming an unsupported statement about the recording or the people it mentions.
6.3 Let a label describe the voice without erasing the quotation
Using WhisperAITranscribe for a public speech is most useful when the transcript preserves speaker continuity, quoted attribution, the source of identification and the limits of the evidence. A presenter can represent many people while remaining the person heard in the recording. The authored case made that distinction inspectable through separate turn and quotation records, including a self-quote and a nested report. Its mechanical checks preserved references and spans; its wrong transformation showed how collapsing roles changes the account. If WhisperAITranscribe provides the initial text, apply the same distinction during source review and later use. The reader should be able to tell who delivered the words at the event, whose words that person says they are representing and what has or has not been independently established. A clear transcript retains those relationships together, allowing quoted voices to remain meaningful without inventing additional people at the microphone.
References: W3C WAI: Transcribing Audio to Text; TEI: Transcriptions of Speech.