How do you translate a recording with several speakers without putting the wrong words in someone’s mouth? The key is to preserve the source of every assertion before you make the target language sound natural. Write who audibly said each line, distinguish a person speaking from a person being quoted, and separate direct speech from reported speech and an editor’s summary. A fluent translated quotation attached to the wrong speaker is a serious meaning error.
People searching for how to translate multi-speaker audio, speaker attribution in transcripts, direct and indirect speech translation, speaker diarization or translating quotes in interviews need a reliable audit trail. The words alone are insufficient. Readers must know whether someone said a statement personally, repeated another person’s claim, read text aloud, denied a quotation, or was merely present while another person spoke. Different languages handle pronouns, quotation marking and reported-speech grammar differently, so the source of a proposition can disappear in a careless rewrite.
This guide shows a practical method for translating conversations, meetings, interviews, panel recordings and quoted remarks. It uses invented teaching dialogue, never alleged quotations from real people, and demonstrates how to mark uncertainty honestly. The method is global and language-neutral: it applies wherever speech must move between languages without shifting identity, responsibility, certainty or the boundaries of a quotation.
Quick answer: track three identities in every quoted statement
Identify the current audible speaker, the person the statement refers to, and the originator of any words being quoted. Those can be three different people. In ‘Mira said that Jo told her the form was ready’, the narrator reports Mira’s account of Jo’s message. The source does not necessarily establish that Jo actually said it, that the form was ready, or that Mira verified it. An accurate translation preserves the layers of attribution instead of converting a second-hand statement into confirmed fact.
Use a working transcript with timestamps and neutral labels such as Speaker A and Speaker B until identities are independently established. Annotate any exact quotation, reported speech, reading aloud, simultaneous speech or uncertain turn. Only after that should you choose natural target-language pronouns and quotation marks. A source record can be imperfect, but the uncertainty should remain visible at precisely the points where the evidence is incomplete.
Four different products that should never be confused
- Verbatim transcript: preserves source wording and meaningful audible details according to a declared convention.
- Edited transcript: improves readability but still represents what people said, with clearly defined omissions.
- Translation: re-expresses the utterances in another language while preserving speakers, claims and relationships.
- Summary: selects and condenses content; it must not be presented as an exact quotation when wording has changed.
A fifth form, interpretation, communicates meaning during the live event. An interpreter may speak in the first person to represent the source speaker, which makes later audio records easy to misunderstand without role labels. A single recording can contain all these modes at once. Label each contribution by function rather than assuming each voice always speaks on its own behalf.
Why attribution errors happen
Conversation is built from rapid turns, gestures, shared context, incomplete sentences and pronouns whose referents are often assumed rather than named. Recording quality can change across speakers. Two voices can sound similar, and a person off camera may be heard but not seen. Automatic diarization systems can split one person into multiple labels or merge two people into one. Translation adds another layer: the target language may require explicit subjects or pronouns that the source leaves unstated.
Once a name has been wrongly attached to a sentence, stylistic editing often makes the result sound even more convincing. A summary can also detach a statement from ‘according to’ and make it appear verified. Quality assurance therefore needs an attribution pass separate from a grammar pass. The question is not just ‘Does the English sound right?’ but ‘Who is responsible for this claim in the original record?’
Step 1: build a source speaker ledger
Start with Speaker A, Speaker B and Speaker C, a timestamp range and an evidence note for each turn. Use stable labels within a session. Do not substitute names from a meeting invitation or camera shot unless the audio or an authorised recording record supports the match. A voice coming from the left channel is not automatically the person shown on the left side of a video. Keep identity confidence separate from text confidence: one may be high while the other is low.
A ledger entry can read ‘Speaker B, 02:14–02:21, says the shipment may arrive Friday; identification unconfirmed.’ If a participant later identifies themself, you can update the ledger after checking the record, but retain the original provenance in a sensitive project. A final publication should not include speculative personal identifiers. Accurate anonymity is preferable to confident misidentification.
Step 2: mark speech acts and quotation boundaries
Distinguish assertion, question, request, instruction, denial, reported speech and quotation. ‘She asked if the package had arrived’ is not the same as ‘She said the package had arrived’. ‘He denied saying it’ is not ‘He said it’. The grammatical reporting verb is an evidential operator: it tells readers what access the current speaker claims to have to another person’s words. Preserve that operator even if sentence order changes.
Quotation boundaries are equally important. In ‘Aisha said, “Do not send it,” and I agreed’, the instruction is attributed to Aisha while the agreement belongs to the narrator. If the target language moves quotations or drops pronouns differently, label the speaker of each segment in working notes before drafting. Do not allow automatic punctuation to decide where one person’s words stop and another person’s begin.
Step 3: resolve pronouns conservatively
Some languages require a subject where others permit omission, or distinguish social roles through pronoun forms. English he, she, they and it can be ambiguous in a long chain of reported speech. If the source does not establish who a pronoun denotes, do not turn it into a proper name for a cleaner sentence. Repeat an established name where it genuinely resolves ambiguity, but mark an unresolved reference if the evidence does not decide.
Consider ‘Mira told Jo that she was late.’ Without additional context, ‘she’ might refer to Mira or Jo. A translation that selects one person silently could blame the wrong individual. An honest translation can preserve the ambiguity or add a brief editor’s note explaining that the referent is unclear. The purpose of editing is to increase comprehension, not fabricate evidence.
Step 4: handle overlap and off-microphone speech
When two people speak at once, represent that overlap according to the declared transcript convention. Do not merge their statements into a new sentence that no single person uttered. An interjection from the audience may be audible but unattributable; label it as an unidentified voice only if the audio supports that a distinct speaker contributed. Background television or a recorded announcement is different from someone speaking in the room.
A live interpreter may simultaneously convey someone’s statement using ‘I’, yet the interpreter is not necessarily the originator of the promise, denial or request. In formal records, role labels make that relationship clear. Keep the source speaker, interpreting voice and translated proposition distinct. The same discipline applies to dubbed media, narrated quotations and passages read aloud.
Step 5: verify names and claims independently
Replay disputed turns at their original timing, compare waveform or channels only as supporting evidence, and consult authorised event records if available. Do not use a language model’s confidence or automatic speaker label as the sole identity source. If a potentially defamatory, legal, medical or employment-sensitive statement is being translated for publication, qualified human review and appropriate authorisation are important. This is a translation guide, not an evidentiary certification process.
Keep a quote audit: original wording, speaker label, claimed source, target translation, uncertainty and editorial changes. A reviewer should be able to see why every quotation mark, reporting verb and named attribution appears in the final version. If the source supports only ‘someone said’, avoid replacing it with a named person’s statement for narrative convenience.
Worked examples: twenty-six ways attribution can shift
The following mini-dialogues are constructed examples. The people named here are fictional examples within the exercises, not factual participants in a real event. Each case isolates an attribution decision and a possible error. When you translate your own recordings, do not reuse these speaker labels as evidence: build labels from that recording’s actual audio and consented context.
1. A speaker reporting another’s claim
Take the invented source record A: ‘Jo says the shipment is ready.’ A is speaking. Jo is the alleged source of information but is not heard confirming it. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Speaker A reported that Jo said the shipment was ready.’ Writing ‘Jo confirmed the shipment’ upgrades a second-hand claim into a verified statement. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Ask whose voice is actually in the recording and whether the alleged source independently appears. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
2. A direct quotation inside a sentence
Take the invented source record A: ‘Mira told me, “Don’t publish it yet.”’ A is the audible narrator; the warning is attributed to Mira as quoted by A. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A reported Mira’s words as “Don’t publish it yet.”’ Putting the entire sentence in Mira’s mouth erases A’s role and the fact of reporting. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Separate the narrator’s framing clause from the embedded quotation before translating. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
3. A denied quotation
Take the invented source record A: ‘I never said the report was final.’ A denies having made a previous assertion. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A denied saying that the report was final.’ Turning the denial into ‘A said the report was final’ reverses the evidential meaning. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Underline never and identify which proposition is denied. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
4. A question reported as a statement
Take the invented source record A: ‘Ben asked whether the order had shipped.’ Ben is reported as asking for information, not asserting that dispatch happened. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A said Ben asked if the order had shipped.’ ‘Ben said the order shipped’ invents a delivery fact absent from the question. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Label the embedded speech act as inquiry and preserve its uncertain answer. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
5. Two possible referents
Take the invented source record A: ‘Mira told Jo that she had missed the train.’ Neither the invented source line nor this context establishes who ‘she’ denotes. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Mira told Jo that she had missed the train [referent unclear].’ Selecting Mira or Jo by intuition can attach a missed journey to the wrong person. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Ask for earlier conversation or leave a transparent ambiguity rather than inventing a name. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
6. A name supplied by a caption tool
Take the invented source record [Voice heard] ‘I don’t agree.’ The recording contains a distinct voice but there is no verified identity. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Unidentified speaker: “I don’t agree.”’ Using a suggested name from automated speaker diarization turns a guess into a public attribution. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Check an authorised participant roster against independent audio evidence before naming. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
7. An off-camera participant
Take the invented source record [Off-camera voice] ‘The door is locked.’ The speaker is audible but is not visible in the frame. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Off-camera voice: “The door is locked.”’ Assigning the line to the person visible on screen confuses video presence with audio origin. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Keep the camera angle and audio evidence distinct in the source notes. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
8. Overlapping turns
Take the invented source record A: ‘We should wait—’ B, overlapping: ‘No, leave now.’ Two people express different positions during an overlap. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A: “We should wait—” / B [overlapping]: “No, leave now.”’ Combining them into ‘We should leave now’ creates a single consensus statement neither speaker made. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Mark the overlap using a consistent format and preserve opposing positions. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
9. A correction of attribution
Take the invented source record A: ‘I said it was late—sorry, Jo said it.’ A explicitly corrects who made the statement. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A initially claimed authorship of the remark, then corrected the attribution to Jo.’ A summary that states only ‘A said it was late’ keeps the retracted attribution. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Check the self-repair and update the final claim while preserving the correction in a transcript. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
10. Quoting text read aloud
Take the invented source record A: ‘The sign says “Entrance closed”.’ A reads an inscription rather than announcing a new rule in A’s own words. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A read the sign: “Entrance closed.”’ Presenting A as the originator of a closure order may misstate authority. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
If the sign’s exact words matter, verify them against the sign rather than only the reading. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
11. Quoted instruction followed by approval
Take the invented source record A: ‘Aisha said “stop the test”, and I agreed.’ The instruction is attributed to Aisha; the act of agreement belongs to A. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A said Aisha had instructed them to stop the test and that A agreed.’ Conflating both clauses into ‘Aisha agreed to stop’ changes the sequence and speakers. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Draw a two-column map of proposition and attributed person. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
12. An interpreter using first person
Take the invented source record Interpreter: ‘I will send it tomorrow,’ conveying Speaker A. The interpreter’s first-person sentence represents Speaker A’s message. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Speaker A, via interpreter: “I will send it tomorrow.”’ Attributing the promise to the interpreter could shift responsibility for the commitment. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Use role labels from the event record and verify whether the interpreting voice was representing or speaking personally. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
13. An interpreter’s own clarification
Take the invented source record Interpreter: ‘Could the speaker repeat the number?’ The interpreter is now making an operational request, not conveying a source speaker’s statement. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Interpreter: “Could the speaker repeat the number?”’ Assigning the request to the principal speaker would distort who needed clarification. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Distinguish interpreting mode from an interpreter’s independent procedural turn. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
14. Reported uncertainty
Take the invented source record A: ‘Jo may have said the keys were here.’ A is uncertain about both the quotation and possibly the remembered content. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A thought Jo might have said that the keys were there.’ Writing ‘Jo said the keys are here’ removes the narrator’s uncertainty and changes deixis. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Preserve may, thought, possibly or similar hedges alongside the reporting relationship. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
15. A nested report
Take the invented source record A: ‘Mira said Jo told her Ben approved it.’ There are multiple embedded layers and no direct quote from Ben. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A reported that Mira said Jo had told her Ben approved it.’ Collapsing the chain to ‘Ben approved it’ converts distant hearsay into direct confirmation. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Write the reporting chain as arrows before simplifying sentence structure. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
16. A group reaction rather than consensus
Take the invented source record [Several people laugh] A: ‘All right, next topic.’ A group produces laughter; A alone initiates the next topic. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘[Group laughter] A: “All right, next topic.”’ Laughter does not demonstrate that everybody agreed with A’s earlier proposal. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Treat audible nonverbal reactions as events, not individual endorsements. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
17. A person reading another’s email
Take the invented source record A: ‘The email says, “Your request is pending.”’ A is quoting a written notice, not making an independent assessment of the request. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A read from the email: “Your request is pending.”’ A summary saying ‘A declared the request pending’ could obscure the source document. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Preserve document attribution and verify current status separately if action depends on it. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
18. A meeting chair’s summary
Take the invented source record Chair: ‘So, I hear that three people oppose it.’ The chair is interpreting a discussion, not obtaining signed votes. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘The chair summarised the discussion as having three opponents.’ Calling this a certified count of opposition exceeds the evidence. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Check the recording for direct expressions of opposition and keep summary versus vote separate. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
19. A name with ambiguous pronouns
Take the invented source record A: ‘Ethan said Ryan called when he was travelling.’ The grammatical referent of ‘he’ is unclear from this standalone line. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A reported that Ethan said Ryan had called while he was travelling [who travelled is unclear].’ A translator may assign the trip to Ryan because his name is closer, but proximity is not proof. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Seek the wider context or preserve ambiguity rather than adding a fabricated travel fact. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
20. A background recorded voice
Take the invented source record [Television] ‘The service is suspended.’ A: ‘Turn that down.’ A recorded broadcast says one line; a person in the room says another. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Broadcast audio: “The service is suspended.” / A: “Turn that down.”’ Merging broadcast and room speech could falsely attribute an announcement to A. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Annotate source channels and avoid recording a television announcer as a meeting attendee. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
21. A request versus a quotation of a request
Take the invented source record A: ‘She told me to wait outside.’ A describes an instruction allegedly given by someone else. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A said she had instructed A to wait outside.’ Do not translate it as A personally instructing listeners to wait outside. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Identify the original addressee, then check whether the source actually names the instructor. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
22. A quote whose negation matters
Take the invented source record A: ‘He said, “I did not sign it”.’ A attributes a denial to another person. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A reported that he said, “I did not sign it.”’ Dropping not converts a denial into an admission, and omitting A’s framing presents it as directly verified. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Audit negation at each level and label whose words are being translated. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
23. A video cut hiding a speaker
Take the invented source record [Cut to reaction shot] ‘We have enough votes.’ The shot changes during the audio and the speaking voice is not definitively identified. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘Unverified speaker: “We have enough votes.”’ Assigning speech to the person on the reaction shot assumes editing and audio are synchronous proof. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Return to the original recording or production log before naming the person. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
24. A paraphrase mistaken for exact speech
Take the invented source record A: ‘Basically, Jo wanted more time.’ A provides an interpretation of Jo’s view, not an exact quotation. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A paraphrased Jo as wanting more time.’ Putting ‘I want more time’ in quotation marks fabricates literal words. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Use paraphrase verbs and remove quotation marks unless exact original wording is known. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
25. Two-language quotation
Take the invented source record A: ‘She said “todavía no”, meaning not yet.’ A quotes Spanish words and supplies a gloss. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘A said she used the phrase “todavía no” (“not yet”).’ Discarding the actual quoted Spanish may lose evidence that those particular words were used. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
If working from audio, verify the quote and spelling; if the speaker is uncertain, keep the attribution qualified. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
26. A proposal mistaken for a decision
Take the invented source record Chair: ‘Ryan suggested that we postpone.’ The chair reports Ryan’s suggestion; no group decision is documented. The first job is not to translate vocabulary but to identify the audible speaker, any reported source of information, and the grammatical act being performed. That distinction prevents a sentence from becoming attached to the wrong person when the target language demands a more explicit subject or smoother word order.
A defensible rendering for this evidence is ‘The chair said Ryan had proposed postponing.’ ‘The meeting agreed to postpone’ fabricates collective assent and changes decision status. Notice that natural prose and faithful attribution are separate requirements. The final text must still show whether a claim is direct observation, a quotation, a recollection, an accusation, a question or somebody’s own interpretation of events.
Look for a later recorded decision before upgrading a suggestion into an outcome. Make a three-column check labelled ‘who speaks’, ‘whose claim or words’, and ‘what is actually asserted’. Then underline every added name, pronoun, reporting verb and quotation mark in the target draft. Each addition should be traceable to the source or identified openly as editorial clarification. When identity is genuinely uncertain, a neutral label is not a weakness: it is an accurate representation of the evidence boundary.
Translation when languages report speech differently
Languages differ in how they mark direct quotations, evidentiality, politeness and the boundaries of embedded clauses. Some permit omitted subjects; others prefer explicit subjects. Some use changes in verb form, quotative particles or reported-speech constructions where English uses ‘said that’. A translator should carry the relationship rather than mechanically copying surface grammar. If the target requires a pronoun that the source omits, choose it only when a referent is supported. Otherwise restructure the sentence to avoid a false assignment.
There is also a difference between the linguistic source of a claim and its truth. ‘Someone told me the file was lost’ reports a statement, not independent evidence of a lost file. ‘The report states that the figure rose’ attributes a claim to a document. ‘The witness denied that it happened’ reports a denial. These are not minor stylistic choices. In institutional contexts, the difference may govern what readers believe has been established.
Working with automatic speaker diarization
Speaker diarization answers the technical question ‘which stretches of audio appear to belong to the same voice?’ It does not by itself authenticate personal identities. Recording equipment, background noise, microphone position, emotional register and overlaps can affect speaker clustering. A tool may confidently label two segments as Speaker 1 when they came from different people. A tool can also split one person’s speech into separate synthetic identities.
Use diarization as a first-pass map. Check changes by replay, compare consistent voice features without overclaiming forensic certainty, and use authorised roster or recording notes for any actual name assignment. Preserve the source of the label in internal metadata. In public educational work, anonymous speaker letters are often sufficient. For sensitive interviews, avoid making participants identifiable through unnecessary contextual details.
If a sentence straddles a speaker boundary in an automated transcript, split it only where the audio shows the turn changes. A punctuation mark proposed by software is not evidence of a speaker switch. The speech chunking guide covers sentence boundaries; this guide adds the separate task of assigning speech and claims to their proper sources.
A quote-review workflow for editors
First compare the final quotation with a preserved source extract. Then record its speaker label, its reported origin if any, the timestamp and the literal meaning of each embedded clause. Confirm negative particles, modality, conditions, tense and reported-speech verbs. In another pass, check whether editorial smoothing has removed ‘apparently’, ‘I think’, ‘someone said’, ‘she asked’ or ‘he denied’. These markers tell readers what sort of knowledge is being claimed.
Next read the translation without source access. Would a reasonable reader understand the utterance as a direct declaration, a reported claim, a suggestion, a verified fact or a denial? Compare that reading with what the source supports. If the target gives greater certainty or stronger authority, revise. A named quotation should have a higher evidentiary threshold than an anonymous paraphrase, not a lower one.
Finally, keep corrections traceable. If an initial speaker label turns out to be wrong before publication, fix the working ledger and the output. If a published item requires a correction, use the publication’s established editorial process rather than silently substituting a new version while leaving misleading extracts in circulation. This preserves accountability without pretending a first draft was flawless.
Practice set: who really said it?
Translate ‘Jo said Mira asked whether the deadline moved’ into one clear sentence, then explain why it does not establish that the deadline changed. Translate ‘I never told Ben to publish’ without turning the denial into a publication instruction. Translate ‘the chair summarised three objections’ while avoiding the word voted unless a vote occurred. Translate ‘we heard a voice from the hallway’ without inventing a participant’s name. Translate ‘the interpreter said I agree’ only after establishing whether the interpreter was conveying a speaker’s view or speaking personally.
After each exercise, ask a peer to draw the attribution chain. If they name a different originator, you may have an unresolved ambiguity or a misleading rendering. Then change one element of the source—the quotation verb, the negation or the speaker label—and check how the target must change. This conditional practice trains attention to responsibility and evidential boundaries rather than merely practising fluent sentence rhythm.
Frequently asked questions
Can I trust the speaker names produced by AI transcription?
Not without corroboration. Automatic tools can separate probable voices, but named identity is a different claim. Check authoritative event evidence where available, and use neutral labels for uncertain participants. Do not publish a sensitive name based on a model’s guess.
When should I use quotation marks?
Use them when the output represents a source’s quoted words according to a clear editorial convention. A translation of direct speech can use quotation marks if its translated status is understood. Do not place paraphrases or inferred intentions in quotation marks as though the speaker uttered those exact target-language words.
What should I do when two people speak at once?
Mark overlap consistently and transcribe what is intelligible from each voice. Do not merge different positions into a consensus sentence. If a fragment cannot be understood, label that segment rather than attributing imagined words to either speaker.
Can I make an ambiguous pronoun explicit?
Yes when independent source context establishes the referent. Repeat a name or role for clarity in that case. If the source leaves the referent ambiguous, an explicit name would be an invented fact; preserve the uncertainty or explain it briefly.
How is this different from translating turn-taking?
Turn-taking describes the floor: who starts, interrupts, overlaps or yields. Attribution and reported speech describe responsibility for statements: who actually said something, who quoted whom and what epistemic status the quotation carries. They interact, but solving one does not automatically solve the other.
Further learning routes
Continue through the Master Art of Translation, How to Translate Easily series index, Sources, Attribution and Verification, and Translation and Meaning Across Languages. The neighbouring guide to interruptions and overlapping speech treats the rhythm of conversation rather than the evidence chain of reported statements.
The durable rule is that every translated claim needs a source, every quotation needs a boundary and every named speaker needs evidence. Preserve those relationships before editing for speed or elegance. Where the recording is genuinely ambiguous, accurate uncertainty serves the reader better than a persuasive but unsupported attribution.