Cite an Instagram Post: What a Transcript Actually Gives You

What a citation needs, and which parts a transcript can help with

Break a post citation into its parts and it becomes obvious where a transcript helps and where it does not. Four of the five parts come from the post; only one comes from the transcript.

• The account that posted it - from the post. The transcript has no author field at all.
• The date - from the post. The response has no date field either; the date printed at the top of this page is the page's date, not the clip's.
• The caption and any on-screen text - from the post. A transcript covers speech only, so anything written on screen is outside it by construction.
• The URL - from the post.
• The exact wording of what was said - this is the one part a transcript genuinely supports, and it is the part people get wrong most often when they quote from memory a day later.

Why you cannot cite a timestamp from the export

So a phrase like "at 00:42" cannot come from this file. It is not a file with slightly wrong times; it is a file with no times. If you need to point a reader at a moment, the timestamp has to be measured off the audio.

Getting a timestamp you can defend

The timings are still in the audio. On a 30-second clip where we know exactly when each sentence is spoken, silence detection found the six boundaries at 0.70, 3.30, 11.22, 14.34, 22.70 and 25.49 seconds, against true values of 0.50, 3.56, 11.00, 14.47, 22.50 and 25.66. Largest miss 0.26 seconds, smallest 0.13.

• Run: ffmpeg -i yourfile.mp3 -af silencedetect=noise=-40dB:d=0.3 -f null -
• It prints silence_start and silence_end pairs. The gaps between them are speech.
• Cut the transcript against those boundaries in a subtitle editor, or read the boundary nearest your quote and cite to the second.
• Cite to the whole second, never to a decimal, and play the moment once before you publish the citation - a 0.26 second margin is inside the second you are naming, not a licence to round further.

Quoting a person: there are no speaker labels

The same applies to anything the model cannot hear: an off-camera voice, a voice note played back on camera, or a translation spoken over the original.

Names, handles and numbers are where a transcript is weakest

We do not publish an accuracy figure, because we have no test set we can stand behind. What we can tell you is what the failure looks like in practice: proper nouns, brand names, handles and figures are the tokens a reader is most likely to check, and they are the tokens a speech model is most likely to normalise into something plausible. That is the worst combination for a citation, because a wrong name reads exactly like a right one.

• Check every proper noun against the caption or the account name before you quote it.
• Check numbers twice; a transcript that says 30 percent where the speaker said 13 percent will not look wrong on the page.
• If a sentence is load-bearing for your argument, play that sentence back rather than trusting the text.

How long a quote actually is, measured

The practical reading: a 90-second Reel is around 590 words of text, which is a long block quotation by most style guides, not a short one. Knowing that before you start cutting is the difference between quoting a sentence and reproducing the whole post.

What the language field does not tell you

In eight uploads of the same English speech, the language field read unknown every time. So it is not a field you can use to support a claim about what language a clip is in. Establish the language from the audio or the post, not from the response.

Source protection, and the one thing we cannot do later

The flip side is that we cannot re-run a clip for you later or hand you a copy next month. Keep your own recording and your own text, and record the date you accessed the post - which is one of the fields a citation needs anyway.