ChatGPT now accepts audio uploads, but interview quotes still need listening

A transcript can sound more certain than the person on the recording. ChatGPT’s October 6 release notes add audio uploads for questions, transcripts, and summaries. For a newsroom, the useful opportunity is finding material in an interview; the dangerous shortcut is publishing the returned wording as a verified quote.

OpenAI’s upload documentation limits audio uploads to paid subscriptions and workspaces, including Enterprise. Availability can depend on settings, region, client, and model. Supported audio includes MP3, WAV, FLAC, AAC and M4A, plus audio-only WebM and MP4. The file must contain a decodable stream; a video file with the same container extension is not equivalent.

A plausible sentence can change the claim

Consider this invented editing example, not output from a ChatGPT test. A speaker says, “We could launch next week if the review is finished.” A working transcript leaves out the condition and renders, “We will launch next week.” Both versions read cleanly. Only one reflects the statement in this fictional recording.

An editor searching for a launch date might find the second version useful precisely because it is decisive. The correction requires returning to the recording, hearing the condition, and checking the surrounding exchange. Fixing grammar or asking the model to “make it more accurate” without that comparison leaves the central problem untouched.

A second risk is attribution. Two people may discuss the same deadline, and a transcript may attach a sentence to the wrong speaker. Verifying the words alone would not catch that. Keep the speaker check separate from the wording check, especially when voices overlap.

From a located passage to a publishable quote

Use the returned notes to locate a candidate passage. Keep the original recording and a reliable position reference so another editor can find it. Listen before deciding whether the passage should be a direct quote, a clearly attributed paraphrase, or omitted because the audio is unclear.

A direct quote needs the actual words. A paraphrase needs the speaker’s intended meaning and any qualification that changes it. If a recording says “could,” do not convert that into “will” in either form. If the statement relies on a previous question, include enough context for a reader to understand its scope.

Resolve names, technical terms, and dates through the recording and relevant source material. Mark uncertainty openly in the working notes. An unanswered identification question is a reason to hold the quote, not a reason to make the attribution look complete.

A successful upload does not prove complete coverage

OpenAI lists a 512 MB audio limit. Longer recordings may be processed in sections when Data Analysis is available, and very long inputs may time out. Its documentation warns that transcripts may contain errors and that speaker identification may be unreliable. Check that the passage you need was actually processed before trusting a whole-recording summary.

Upload only material you are authorized to process under your consent and workspace rules. This is private source work: a citation to an uploaded interview does not make that recording a public-web source or establish search visibility. Search Engine Answer has not run an audio accuracy or speaker-identification test for this report.

The evidence-led publishing guide separates a supported statement from an editorial inference. Use the generative AI content audit checklist before publishing, particularly when model-assisted extraction supplies a quote, number, or attribution.