Choose between transcripts and summaries by future evidence needs, permissions and review effort, with a storage example that shows where the real costs sit.
Direct answer: Keep a short source-linked summary for routine discovery, and retain a full transcript when you are authorised to do so and expect to verify quotations, revisit context or support text-based access. Do not discard the route to the original recording just because a summary reads well. If your permissions or retention rules do not allow the transcript, keep only the permitted reference information and state the resulting limitation.
A full transcript and a summary solve different problems. The transcript preserves a much larger portion of the spoken record, while the summary selects what seemed important to a particular reader at a particular time. Neither automatically contains everything visible in a video.
The decision is therefore about future evidence, access and handling obligations, not just file size. An archive that is cheap to store but impossible to check is not useful research infrastructure.
Make the retained-evidence choice
The retained-evidence choice is an editorial method that asks what a future user must be able to recover. Identify the likely task first: finding a relevant source, checking exact wording, comparing answers across interviews, or understanding material without listening.
This extends building a searchable media knowledge system by deciding how much of the underlying evidence to retain. It does not prescribe a universal retention period or create permission to copy material.
Before collecting more files, check the agreement, licence, institutional policy and storage arrangement applying to each source. Use approved access controls for private interviews. A public video is not automatically available for unrestricted transcription and redistribution, and an interviewee's permission for one project may not cover every future use.
In the UK, copyright exceptions have specific conditions; the government's guidance does not treat private study as blanket permission to copy whole works. For a consequential reuse decision, obtain qualified local advice. Requirements vary by country and context. UK guidance on copyright exceptions.
Compare the records by what they preserve
| Record | Useful for | Important limitation |
|---|---|---|
| Source-linked summary | Finding the relevant source and its main contribution | Omits material according to the summariser's judgement |
| Full transcript | Searching wording and reviewing a wider verbal context | Can contain errors and omit visual information |
| Summary plus transcript and recording reference | Discovery followed by deeper verification | Requires more checking and disciplined access management |
My default recommendation is a brief summary linked to a permitted transcript when the source is important enough to revisit. For disposable background reading, a summary and original link may be sufficient. For sensitive material, keeping less can be the correct outcome even when more would be convenient.
Do not label a summary “transcript”. A reader may reasonably expect a transcript to represent the speech, whereas a summary intentionally selects and paraphrases. Similarly, mark an uncorrected machine transcript as a draft rather than letting its length imply accuracy.
Keep access needs distinct from personal preference
Text can be the primary route into audio material for some readers. W3C distinguishes basic transcripts, which include relevant speech and non-speech audio information, from descriptive transcripts that also capture necessary visual information. A short summary does not necessarily provide either. W3C guidance on transcripts.
Ask who will use the archive and what they need to understand. Do not delete a fuller text record merely because you personally prefer listening. Equally, do not assume that a word-only transcript makes a visually explained process accessible without descriptions of its important on-screen actions.
This article is an archive decision method, not a complete accessibility compliance assessment. If you publish media, assess the requirements for that publication separately. Your private research summary and the text alternative offered to an audience may have different purposes.
Preserve the relationship between versions
Give the source, transcript and summary related identifiers. Record the recording version, transcript status, creator, date and relevant timestamps. If you correct a speaker label or quotation, update the dependent summary where the change matters.
Keep the original transcript separate from a corrected working copy when the project requires an audit trail. Do not silently overwrite the only evidence of what an automated system produced. However, follow applicable retention instructions rather than retaining every intermediate indefinitely.
For each summary claim, include a locator that leads to the supporting passage. “Useful discussion of reliability” may help discovery, but “explains the recovery limitation at 24:10” helps verification. Mark your interpretation separately from what the speaker states.
Work through the cost of 22 transcripts
Imagine a researcher with 22 interviews. The sizes and times below are illustrative assumptions, not typical file-size measurements or product specifications.
Assume each plain-text transcript occupies 180 KB and each summary 12 KB. The transcript collection uses 22 × 180 = 3,960 KB. The summaries use 22 × 12 = 264 KB. With decimal units, that is 3.96 MB versus 0.264 MB, a difference of 3.696 MB.
That difference is modest in this scenario. It would be misleading to present deleting transcripts as a major storage saving without checking actual files. Audio and video may be much larger, but their size is a separate calculation.
Now assume checking each transcript sufficiently for the planned use takes 18 minutes, and producing a source-linked summary takes six. Transcript checking costs 22 × 18 = 396 minutes. Summaries add 22 × 6 = 132 minutes. Add 22 minutes to inspect identifiers and access settings: combined effort is 396 + 132 + 22 = 550 minutes, or nine hours ten minutes.
Those times do not guarantee complete transcription accuracy. They illustrate why review and maintenance can dominate storage. If only four interviews are likely to support quotations, prioritise fuller checking there while clearly labelling the status of the others. Do not imply that a partial review verified every line.
Test the archive with a future question
Choose a question that was not the focus of the original summary. Can you find whether the speaker mentioned an exception, identify who said it and return to the relevant recording? A summary-only collection may fail because the detail was deliberately omitted.
If the transcript answers the search but its wording is uncertain, use the recording to check. If the recording is unavailable, retain the limitation. A long text file does not become an unquestionable original merely because it is the most complete surviving record.
For a small archive, run this test manually. Buying a search system before establishing what your records preserve can make missing evidence easier to overlook rather than easier to recover.
Review one collection in 45 minutes
- Spend ten minutes listing the future tasks and the permissions governing the collection.
- Use 15 minutes to inspect three representative records, including a sensitive or difficult example.
- Spend ten minutes testing a question that the summaries do not already answer.
- Use the final ten minutes to choose the retained record for each source type and note who owns future corrections.
Do not delete the fuller material until you have confirmed that removal is permitted and that the remaining record supports the intended task. If permissions are unclear, stop expanding or redistributing the archive and resolve the handling question first.
Related guides
Frequently asked questions
Can a good AI summary replace the original transcript?
Only for a task that genuinely needs a summary rather than recoverable wording and context. A fluent summary may omit an exception or a detail that becomes important later, even when its main points are accurate. Keep its source locators and label it as a summary. If you need to quote, compare statements or investigate a disputed attribution, retain the permitted transcript and recording reference. Before submitting private material to a summarisation service, check authorised handling. A shorter document is a useful discovery layer, not evidence that everything worth retaining has been preserved.
Should I keep uncorrected transcripts at all?
Sometimes, as provisional search aids with clear labels and a route to the recording. They can help locate candidate passages even when every word has not been checked. Do not use an uncorrected transcript as a verified quotation source or let it circulate without its status. Record which sections have been reviewed and which remain uncertain. If errors are so frequent that searching becomes misleading, the file may not justify keeping. Follow retention and permission requirements in either case. The choice is not between pretending the draft is accurate and deleting it immediately; limited, labelled use can be appropriate.
Does a transcript capture everything in a video?
No. A word-focused transcript can omit diagrams, demonstrations, gestures and other visual information necessary to understand the content. Check whether the archive needs descriptions of those elements or a direct reference to the relevant frames. A summary may also omit important sounds or speaker changes. If text is someone's primary way to access the material, assess those needs explicitly rather than assuming any text file is equivalent to the video. For a narrow research question about spoken wording, a basic transcript may suffice, provided you do not generalise that adequacy to every possible use of the recording.
How should I handle a source that later disappears online?
Keep the identity, original link, version information and date of your last verification. Use any copy you were authorised to retain according to its terms, but do not acquire or share an unauthorised replacement merely because the public link stopped working. If you only have a summary, acknowledge that exact wording and context may no longer be checkable. For a claim central to new work, find another suitable source or narrow the statement. Disappearance is a reason to reassess the evidence path, not to upgrade your old paraphrase into a definitive transcript.
Should I use one retention period for the whole archive?
Not automatically. Different sources can have different permissions, sensitivities and research purposes. A public reference note, a confidential interview and a commissioned recording may require separate handling decisions. Establish the relevant project or organisational policy and record the reason for retention, rather than choosing a duration because storage is cheap. This article does not prescribe legal retention periods. Requirements vary by jurisdiction and context, so seek qualified advice where necessary. A consistent review process is useful, but consistency should not mean ignoring material differences between the records you are responsible for keeping.
How do I avoid correcting the same mistake in several files?
Keep one clearly identified corrected transcript and link summaries to its relevant passages. Record a meaningful correction once, then inspect the dependent summaries for claims affected by it. Avoid copying large transcript sections into many unrelated notes, because each copy becomes another maintenance obligation. Preserve original versions where an audit trail is required, but mark which version should be used for current work. If your existing tools cannot maintain sophisticated links, a shared source identifier and correction note can still work. The objective is traceable dependence, not a complex database that nobody updates.
Sources and verification
- W3C: transcripts, consulted 11 September 2026 for the distinction between basic and descriptive transcripts.
- UK Intellectual Property Office: exceptions to copyright, checked for the limited, conditional nature of research and study exceptions. This is not case-specific legal advice.
- The parent was read locally because its public route was unavailable. The retained-evidence method and file-size example are editorial assumptions.
This article is practical guidance. Apply it in proportion to your tools, evidence, risks, and responsibilities.
