A local archive publishes a collection record in English for a conversation conducted mostly in another language. The record's only language field says English. A listener opens the audio expecting English speech. A second visitor wants to find interviews in the language actually spoken and misses the record entirely. A third downloads a file marked transcript and assumes it contains the words as recorded, although it is an abridged English rendering. All three interactions can be wrong without any obvious broken link.
The archive did not necessarily mishear the audio or mistranslate a line. It merged several entirely different attributes: language of the recording, occasional alternate speech in it, language of the catalogue description, and languages and forms of its related files. Repairing this confusion does not require making every recording available or creating transcripts for every part. It requires a language note small enough to maintain and specific enough that “English” is never the whole answer unless English is all the evidence warrants. This method is about describing identified representations, not changing anyone's access agreement or putting recordings into new services.
Start with the record the listener is about to open
Write one simple sentence about the audio as audio. Use phrasing like “Interview mainly in Spanish, with questions in English and occasional English replies,” or “Interview in the recorded language; some segment language still being reviewed.” Choose terms you can support on the evidence you have and don't fill a field from someone's name, accent or residence. If contributors used a preferred self-description for the language, preserve it where relevant, optionally with the catalog's documented controlled term for access. Avoid treating a descriptor as evidence about a speaker's identity, background or citizenship.
Next list attached representations: verbatim transcript, edited transcript, translation, subtitles, rough working notes, title and summary. Name the language of each and the kind of relationship it has to the recording. “English summary” doesn't mean the whole interview exists in English text. “Transcript in Spanish with intermittent English preserved as spoken” doesn't mean English subtitles are available. A short note for users can describe those differences without including every metadata field. The detailed ledger is there to make the useful public sentence accurate.
The Library of Congress MARC 21 language code documentation is one example of a vocabulary capable of identifying different language roles. Among other distinctions, it addresses original-language relationships and languages found in accompanying forms, and includes an oral history example. The companion language note field carries textual descriptions. A small archive need not implement MARC to use this basic idea. A coded search field enables consistent retrieval; a display note tells a human what the code refers to. Neither should be tasked with doing all the work alone.
Make ambiguity an explicit state. The team may lack a reviewer for one short excerpt. A remembered song fragment or personal name might be in another language without constituting a second sustained spoken section. A short alternating exchange can be well documented even where the project doesn't know how it wants to classify it across the entire collection. Distinguish unverified, mixed with documented spans, and not yet described. Those states are more informative than temporarily tagging every item with the language of the cataloguing interface.
Record changes at a scale a review team can actually maintain
For a recording with several longer stretches, use small time-based notes rather than assigning another complete language label to the whole session. A note could track where a narrated section begins, where the interviewer switches, or where the recording has a separately introduced reading. Clarify whether these are rough locators or carefully verified segment bounds. They can help an editor find content and help a listener decide whether to use the transcript. They do not always need to be exact to a frame, particularly where two people overlap or a speaker switches mid-sentence.
Use a start and end time within the exact version consulted. An edited interview and its longer source copy can have different time offsets; a note in one cannot be pasted into the other as if the timelines were identical. Include the version ID or a stable filename and indicate where rounding was applied. The visible cue can be brief: “after the second question about the market,” or “opening two turns.” A human clue is valuable where a platform retimes playback or changes the way leading silence is handled. Keep the time marker when it can be trusted, and let the clue be a second route rather than replacing it.
Try several scope labels before deciding the project's segmentation rule. A block may involve continuous speech in one language, conversation moving among two, a short recited fragment, or partly unidentified speech. If the aim is discovery for researchers, use an explicit policy for which cases count under each filter. If the aim is editing and locating points for review, finer spans could be worth recording. Do not create a token-by-token tagging workload merely because one research assistant has a long afternoon. A practical note balances descriptive usefulness and the team's ability to verify and preserve it.
The Oral History Association's archival principles and practices discusses transcripts, captions and representing interview collections with suitable metadata and repository standards. It is useful background for keeping versions legible rather than trying to express all context in one description. A team can preserve a basic descriptive note when no transcript exists. Never describe an automatically produced draft as reviewed or publish-ready merely because its language code looks plausible.
Keep speech, translation and description from silently replacing each other
A transcript can preserve spoken words, but transcription always involves choices about how to represent starts, pauses, speakers, punctuation and difficult passages. A translation represents the content in another language and can also be shortened, adapted or annotated. For discovery, both may be useful. Their usefulness is not a reason to advertise a translation as the original. Name the relationship precisely: full original-language transcript, translated extract covering specified sections, bilingual caption file for a short exhibit item, or summary only. If the recording can be listened to in the original but no readable transcript is available, say so.
Take the version as seriously as the language. A working text file can reflect an early cut of the audio; an accessibility caption may be generated for a shorter approved version; a later transcript can correct an earlier mistaken speaker boundary. A form named final tells a reader none of that. Record the identified media relationship, version or date, transformation type, and review status. Do not require someone opening a public file to infer by comparing the first two lines whether it is a full transcript or a translated selection. Put those facts in the catalogue note and in the file where feasible.
Maintain explicit alignment where it is useful and sufficiently stable. One translation might retain paragraph numbers from the original-language transcript. A separate caption track may use media time markers. Short quoted sections can reference a start and end plus a position clue. If the translator deliberately reorders paragraphs or combines material thematically, make that editorial handling visible rather than pretending there is a one-to-one match. An archive can share a summary as a summary without committing to provide an exact written version of everything the person said.
Avoid describing differences with dismissive language. If a translator offers two plausible English versions of a term, a note can say that it carries more than one sense and name the applicable context without claiming that one wording is defective. If the underlying recording is unclear, mark the passage as uncertain rather than hiding the uncertainty by translating a guessed word fluently. An account can also use quoted language for effect or repeat a phrase from someone else. Identify the role of an embedded utterance when that matters. A bilingual turn does not mean each speaker consistently used one language throughout.
The published label can use a plain sentence: “Audio primarily in Portuguese. English introduction, short Portuguese transcript for the first section, English summary.” If the project has an appropriate review process, give each claim a checked state before it goes into public copy. Factual honesty about a missing representation is part of accessibility. A listener should be able to decide whether the material is useful without opening three files and inferring that none contains a full original-language transcript.
Synthetic example: three windows into the Pine Market interview
All participants, programme, documents and language observations in this example are invented. A market-history project interviews a former vendor. She tells the main account in Portuguese; a host gives a brief English introduction and asks some English follow-ups. Midway, the narrator reads a short portion of a Spanish-language notice posted on a market wall in the story. A project assistant writes an English description, and a translator produces English paragraphs covering only the portion of the recording about delivery mornings. A site template initially labels the recording English because it imports the description's language and displays the English paragraphs under Transcript.
The first fix is identifying the actual objects. Item M7-A in the fictional project is the source interview. Its language note says “Main account in Portuguese; English introduction and selected questions; includes a short reading in Spanish.” The description is M7-D and is labeled “English catalogue description.” The translation is M7-X1 and is identified as a selective English translation of a named segment. The host's introduction is not silently dropped from the audio note, but it also does not replace the account's primary use of Portuguese. The record says no full transcript at this point. The site now offers accurate expectations even before any additional editorial work is possible.
The volunteer adds three time markers for help finding the representative stretches. One gives the beginning of the account after the English framing. The second marks the period where the host's questions are heard more often. The third identifies the short reading from a posted notice. The times refer to the source version and are approximate. No speaker is assumed to identify with a nationality because of their use of a language, and no inference about the vendor's background enters the catalog because it isn't needed to locate the interview. A language note describes speech at a particular place in an object. It doesn't label the person.
Two months later an exhibition uses a shorter reviewed clip with no introduction and only one English question. The clip contains Portuguese speech but not the Spanish item referenced in the full interview. Its own note should reflect only what the clip contains: Portuguese account with one short English exchange. Clicking from a search result for a Spanish reading to that clip would be misleading if there is no way to find or identify the reading from the result. If the site offers related results, it can display the source record and describe the clip as a related excerpt, rather than claiming the excerpt contains Spanish. Availability and navigation remain separate matters from speaking accurately about the item.
Now a second editor offers a bilingual transcript but uses the older exhibit clip's timings. That document is a useful candidate only if displayed with matching media and labeled for the accurate scope. The older selective translation of the source and bilingual clip transcript are separate representations. Neither overwrites the other merely because both contain English. The catalogue entry can link both when appropriate and label them by content, scope and version. A note tells readers that the introduction and reading occur only in the full source. With those roles distinct, a small collection gains meaningful descriptive depth without creating unverified content.
Describe a language switch without inventing a speaker's reasons
People can shift language mid-answer, repeat a phrase, draw from a document, use a borrowed technical term or address another person directly. The recording shows the words and provides context; it rarely establishes all the reasons for a shift. A catalogue does not need to infer what a switch says about fluency, audience, upbringing or intention. Record the language when responsibly identified and the visible or audible scope that affects retrieval or use. If a switch matters to the historical interpretation, that is a separate editorial claim that needs evidence beyond a lone language note.
Develop one or two local conventions. A collection might record that selected passages contain sustained conversation in another language; named quoted fragments are separately described in segment notes but do not imply a whole alternate transcript; recurring greetings alone do not automatically populate every facet. Apply the conventions consciously and allow reviewer notes for exceptions. A broad search can offer “contains or discusses” if the distinction is explicit. Avoid a filter name that promises audio predominantly in a language while its logic includes all items with a single related English field. These distinctions are about making user expectations match actual implementation.
If listeners disagree about the language of a brief passage, record the disagreement respectfully and get appropriate review. A machine's confident guess or a caption based on perception of someone's surname cannot settle it. Do not take a recording to a new outside service merely because the necessary reviewer isn't immediately available. Keep the language field tentative and let the rest of the documented item remain findable. A little narrower search coverage while clarification is sought is better than giving many visitors detailed-looking metadata with no reliable basis.
Sometimes a passage is transcribed into a different script from the one a contributor expects. Script is another attribute, not necessarily another language. If the catalog records script for stored text, separate source spelling, transcription form and any romanized search aid. A speech recording itself doesn't make a written script choice; its two related transcripts can choose different writing systems without implying an alternate spoken recording. Preserve literal language labels and translation relations without deriving identity from a UI's language menu.
Make site search ask the question the reader means
A practical browse interface could offer distinct facets for spoken content, subtitles, full transcripts and summary language. Not every catalogue needs four visible controls. The underlying model should at least let staff tell these roles apart. A collection with only one browse control can label it “language spoken in audio” and list other materials under a separate description. A user wanting English text can still search titles and notes, but the website should not claim that these results are all English recordings. The language of the interface and its text indexing can be chosen separately from the material users are browsing.
Check two sorts of results. Specific lookup should point to an actual representation matching the chosen facet; selecting “has a Spanish transcript” ought not to return an English transcript with a short untranslated Spanish quotation unless the collection transparently defines it that way. Broad discovery can find material connected to a language without implying equivalence. A label saying “English description available; audio in Portuguese” can still help an English reader find relevant research material. A label saying “English interview” would be wrong unless the audio description earns it. Let people understand why a valid result showed up.
Make the response useful when search results are empty. In some collections, no language-aware filters have yet been applied to older records. A no-result message should not represent the absence of interviews in that language, especially where description is incomplete. Explain whether search only covers verified coded records or the entire compatible catalog. Offer a non-identifying route to browse incompletely described records when policy permits it. Do not surface restricted material to demonstrate broad coverage. Catalog metadata, item access and content permission are separate questions.
Where a repository exchanges MARC records or exports adapted metadata, follow the standards appropriate to that catalog type. The Library's 041 guidance and 546 guidance provide definitions and examples, not a requirement to improvise encoding by guessing which subfield sounds right. Have relevant cataloguing staff check how they want to express particular cases. A lightweight website can model plain roles first, then map to the institutional format with documented choices. Do not pick a code solely because its display label matches a word a volunteer recognized.
Synthetic example: the link that sounded bilingual only on the page
This shorter scenario is fictional. An eight-minute exhibit player shows a mostly Catalan account and opens a caption menu offering an English track. The archived unedited material also contains a longer Spanish exchange late in the session. A migration script takes the three language labels in the collection-level note and adds them as though all were simultaneously present in the exhibit player. During testing, Spanish discovery begins returning the eight-minute clip even to viewers who could not locate such a stretch there. English-audio search also acquires the item because English captioning was mistaken for English sound.
The team does not delete language data to fix the mismatch. It creates a correctly described relation: archive source includes the sustained Spanish exchange; clip uses Catalan speech; select English caption track applies to the clip. A discovery result for documented Spanish speech can show the appropriate source-level record, subject to established access arrangements. A broad catalogue entry can show both representations and state which contains what. The migration now preserves search, position and meaning rather than making a broad description masquerade as detailed player metadata.
A reference librarian then opens an older transcript labeled Catalan in the working system but finds that it covers the whole original recording, including the Spanish section as spoken. The full label becomes “original-language transcript of source recording, including Catalan and Spanish segments,” not “Catalan translation.” This is a correction of media description, not a reassessment of who may view it or an act of publishing the working file. Authorization and appropriate handling still apply. A case like this is a useful acceptance test because it distinguishes a whole work, an excerpt, a related textual item and a caption with separate language roles.
Audit rendering and placement, not just code assignments
Open the actual user facing display for every representative case in a test set. Check the requested view, the title, the exact linked representation and the language note near the action it explains. A record can have correct metadata but use a generic play button with a confusing language chip elsewhere on the screen. A file labelled only .txt may become harder to identify after download if it lacks the descriptive note visible online. Ensure it remains clear whether the person selected a translation, transcript, captions or an original recording appropriate to their permitted use. Accessibility affordances work better if their scope is obvious before someone clicks.
For text, confirm that the expected scripts, accented letters and direction display in the environments you actually support. A working JSON import doesn't prove the print export renders the text properly. One rendering issue should not silently become a content change. If a fallback version uses different script or removes styling, label the output and retain the reviewed source form. A system should not replace original speech metadata because its UI default language controls headings, nor treat a translator note as talk occurring during playback. Read a few actual affected entries, not only a dashboard of successful conversions.
For captions, check which audio version they accompany. Even a well translated track can become hard to use when shifted several seconds compared with its real clip. That is an alignment issue, not necessarily a bad translation. Verify timing using known markers, identify the track and player versions and separate quality from language metadata. If a particular caption is unreviewed, note the status rather than hiding the limitation under a confident designation. A user can handle some uncertainty if they know where it lies, and a future reviewer will know exactly which evidence needs checking.
Build a version map that can preserve a missing relation
Begin with arrows for actual derivations. The source recording might lead to an original-language transcript, which leads to a selective translation. A caption file can be based on a different excerpt, perhaps with revised subtitle breaks and no translation in portions where the spoken language already matches the selected captions. Draw from source to derivative and label the relationship. Where the editor doesn't know which audio version produced a transcript, leave the arrow unresolved rather than pointing it to the newest recording as a default. A reusable description distinguishes “derived from this version” from “about the same person and topic.”
A simple version map can carry record ID, file role, language, coverage and timing basis. Consider a transcript A that follows the direct recording, a translation B of transcript A through chapter three, and a second transcript C created for a demonstration clip from a different cut. A download screen can make B and C both appear to be shortened translations. The version map lets you rewrite two short labels, “English translation through chapter three” and “original-language transcript of demonstration clip,” without hand-auditing every line each time another interface is built. If the source of a version remains unknown, don't offer exact timing links until they can be checked.
Some documents are not derivatives in the same way. An English catalogue summary may use interviews, records and editorial selections by the archive team. It doesn't necessarily correspond to a contiguous passage in the source. A community review note in another language can describe the record without representing the original speaker's words. Label separately whether the text quotes, transcribes, translates, condenses or discusses the recording. This supports citation, helps a reader choose what to open and prevents an institutional voice from being silently attributed to a narrator.
Audit a copy that travels without its player. A caption export could retain a header but lose the collection page telling readers what version it accompanies. Include a small language and scope note in the file or in approved accompanying documentation. Consider how people might read the result months later: does unqualified text look like a full translation? Is there enough of a source pointer to find the right edition? Working context that travels with the version often prevents more trouble than a meticulous private schema whose useful fields never leave the catalogue admin screen.
Make mixed-language work understandable without claiming every word has one home
Some stretches will not fit a simple block model. People can mix linguistic forms within a turn, and the description intended for a website may be intentionally broader than a linguistic analysis. Choose whether interval markers indicate the presence of a language, an approximate passage in which it predominates, or a field with an uncertain beginning or end. Communicate those definitions to the people building search and citation. A short recognized phrase nested in a different narrative doesn't imply a second full recorded retelling. If the people responsible for a collection choose to tag that phrase for discovery, name it as a quoted or embedded piece when appropriate.
Plan for simultaneous voices and background audio. An interviewer can ask in one language while another person's immediate response starts before the sentence ends. A radio or song audible in the room may not belong to the spoken interview. Unless it is germane and responsibly described, the language note can focus on foreground content and treat background identifications as separate observations. A timeline should not imply that one code has replaced another at a perfectly crisp time if the source is overlapping. A brief label can say “overlapping questions and answers” without pretending the recording contains cleanly sliced blocks.
The note is also not a judgment about who qualified as an expert. Invite reviewer corrections to exact terms and definitions in a way that can be logged without making private biographical inference public. Record source and method more concretely than “AI says” or “sounds foreign.” The archive might say “cataloger's identification reviewed against the transcript by an appropriately experienced speaker” or simply “unidentified speech at 08:14; not used for search facets.” A small collection should not have to supply fine-grained claims it cannot verify. Its deliberate unknowns give future catalogers something more useful than overconfident inherited tags.
Publishing verbatim language judgments can carry special issues where an excerpt includes family, location or sensitive historical content. Keep the methodological note distinct from the actual words and timestamps shared. If there is no authorized public reference to a disputed passage, it is still possible to document general scope privately and publish a broader truthful note where appropriate. This article's procedures describe metadata; they do not give mechanical permission to reveal sensitive passages in order to prove a label.
Test imports with counterexamples people routinely miss
Select a small set of known materials representing different relationships: mostly one spoken language with a short introduction in another; a translation whose original source contains two languages; an original-language transcript with embedded quotations; a record with English catalogue metadata and no English text or speech. For each item, describe the expected display and retrieval outcomes before import. Then check them after import as a normal authorized viewer rather than expecting internal database values to prove meaningful public behavior. Two files with the same language can legitimately appear in different facets because they perform different roles.
Try the reverse direction as well. Start from the user-visible record and ask a colleague to identify which exact version meets a stated need. “Find the interview with an original-language transcript of the first part” is a different task from “find a translation of the whole recording.” The test can reveal whether the catalogue note is buried where a user cannot find it, whether the supplied links open the exact artifact, and whether absence of a subtitle is accurately disclosed. A good outcome includes knowing when the catalogue has no material meeting that test, rather than finding a vaguely related record under a broad search.
Correct a role mapping at its source rather than editing one page. If the web display inherits an overall language tag from the summary field, the problem might affect every imported record, not just the visible example. Trace the mapping with the responsible staff and preserve the earlier record where needed for reconciliation. In a controlled QA fixture, change only the summary-language field and confirm that the speech filter remains unchanged. Then change the spoken metadata in a compatible fixture and check the intended speech filter. Tests like these make the different fields observable; they don't require moving confidential interview files into a testing system.
When advising a small project, prioritize mislabeling that changes whether an item is discoverable or what a visitor thinks they will hear. A punctuation difference in one note may matter less than an import silently dropping a second spoken language for dozens of interviews. Estimate actual scope if there is lawful and appropriate access to that data; otherwise use a sample and mark uncertainty. A corrected example is evidence of that correction, not a count of how many other records were repaired. Document representative findings and hand off specific tests rather than issuing a broad guarantee.
Choose a useful public note even if two formats won't be created
Editors can be tempted to promise a full transcript, translation, subtitling and review whenever an interview includes language variation. Some projects cannot promise that. An honest catalogue can still make progress with a labeled audio record, descriptive note, indication that selected passages use another language, and a statement that no parallel transcript is present. This approach improves discovery without quietly converting a broad intention into a commitment. The reader is better served by knowing a text is absent than spending time opening a nonmatching summary.
A catalogue note should be based on available evidence, clearly phrased, and small enough that changes are actually maintained. Standard template phrases can help: “Spoken content,” “Original-language transcript,” “Translation,” “Partial summary,” and “Clip captions.” Wherever appropriate, replace broad template language with a stable scope and version clue. If there are no documented alternate versions, keep the line concise and skip blank claims. People don't need a cumbersome table for a single clear recording, but they do need details about different roles when those roles affect use.
Share a successful distinction with future volunteers rather than a list of failed labels. An example saying “The item appeared in an English-text search because of a partial translation, not because it is English audio” can teach as much as a formal mapping chart. Keep it based on appropriately cleared training data or an invented case. One transparent chain from the available speech through the record to the related transcript can improve future cataloguing. That is the point of a language note: help people find the represented material without having to infer what a word in a form was supposed to mean.
Save a minimal, portable language note
A useful internal record gives item ID, specific media or text version, kind of representation, which languages were identified and by what source, segment locators where useful, any uncertainty, and current evidence. A public summary needs less detail but should include the key distinction: spoken language, available text and basic scope. If the record is linked to an exhibit or external library exchange, preserve a reviewed mapping of what the downstream label actually claims. No export should quietly fill missing fields with the website's interface language.
Time-stamp a real correction. A later reviewer might identify a short segment previously marked uncertain or recognize that a supposedly full translated transcript stops halfway. Update the affected role and description rather than announcing the whole project misidentified all audio. Keep a brief record of what changed and why. Ask whether a feed or embedded record still carries the earlier claim. A corrected source record doesn't automatically verify every dependent surface. Maintain these checks with internal identifiers and avoid publishing unnecessary personal information while explaining a language metadata change.
Print the language note once on a mock reference slip, then detach it from the catalogue page. Ask a colleague to tell you which version it describes and whether another version holds additional language material. If they cannot tell, add the specific medium, duration or coverage and a persistent internal reference suitable to the audience. A helpful fallback is a human-readable phrase such as “eight-minute exhibit clip; separate full recording in collection” rather than an opaque sequence number with no context. Test the attached file in the same way. A caption can have a perfectly correct internal language code but fail this handoff if nobody knows which recording it belongs with. This exercise also exposes summary wording so broad that readers assume any available text is a transcript. A one-line label that travels with the asset prevents the detached file from teaching the wrong expectation.
A strong handoff lets one reader answer both “Can I hear this language here?” and “Is the text I am about to open a translation, a transcript or a summary?” At the same time, another reviewer should be able to point to the exact version that gave that answer. The purpose is not a comprehensive statement about everything a participant said. The purpose is making clear how to find and understand the forms an archive actually describes. Those distinctions are small enough to keep and consequential enough to deserve their own note.
Editorial note: Drafted with AI assistance for Apiary. The example interviews, records, language counts and editorial results are synthetic.