What this is. This page embeds the White House's "TRUMP TV: The Essentials Station" stream and, alongside it, a transcript generated continuously from the same feed: speech is transcribed and split into sentences, each sentence is attributed to a speaker by voice, and speakers are named where a voice has been identified. Labels such as Speaker 3 are voices that have been told apart but not yet named, and a line marked unattributed was too short a clip to place against a voice at all. Where the feed switches to a language the speech model does not cover, the passage is marked as not transcribed rather than rendered: fed a language outside its range the model does not fail, it produces confident phonetics that read exactly like a poor transcript of English, and a gap you can see is better than text you cannot check. A line is shown under a person's name only when that line itself matches the voice strongly enough; where it was merely grouped with the speech around it, the neutral Speaker label is shown instead, because grouping a line and asserting who said it are different claims and only the first is safe on a short clip. Beside the faces, a general detector labels what else is in shot, so a cut to vehicles or aircraft is recorded rather than read as an empty frame; those labels are generic by design, so a military truck is simply a truck. A name shown in italics with a tentative match tag cleared the acceptance bar only narrowly. The transcript trails the video by a short delay and is machine-generated, so treat it as a finding aid, not a record of what was said. Collected for academic study; the archive runs from the stream's first second.
How it came about. On Friday 18 September 2026 the White House barred CNN, MS NOW and Politico from the press pool, and their reporters' credentials were deactivated, in some cases collected at the gate the next morning. Other outlets, Fox News among them, then declined their own pool duties, and the television pool stopped supplying shared coverage of the president's events. The three barred organisations sued the administration on Monday 21 September, arguing that revoking access over their reporting violates the First Amendment. That same evening the White House began streaming its own channel. The first segment this archive holds is timestamped 23:00:21 UTC on 21 September, which is 7:00 PM Eastern, the hour the launch had been announced for.
What the system sees. Alongside the transcript, faces on screen are detected and boxed on the exact frame they were measured from, and a green box marks the person the audio says is speaking. A name is only attached when the face matches an identity-labelled reference photo, from Wikipedia or Wikimedia, above a fixed similarity with a clear margin over the runner-up. Open evidence under any person to see the frame, the score, the reference photo used, and the alternatives that lost. Faces with no confirmed match stay unnamed rather than guessed at. Most people on this feed have no findable photograph, so a second route exists: where someone states their own name aloud, that name attaches to their voice, and the voice can then carry it back to the face it speaks with. That is a weaker claim than a photograph, since it says the face belongs to the voice that gave the name rather than that the face was verified against a picture, and it is shown as a tentative match. It only applies where the voice holds most of that face's screen time, clearly beats the runner-up, and that face is a real share of the voice's own time, which is what stops every listening face in a rally crowd being labelled with the speaker's name.
Where this is from. The channel replays speeches and events without labelling them, so a passage can be hours or years old with nothing on screen to say so. The transcript is therefore cut into segments at long silences, and each segment is put to a search agent that reads what was said, who was on screen and any text in frame, and proposes the occasion. Nothing is stated on that basis alone: a verbatim line from the segment is then searched, and the occasion is only shown as confirmed when an independent page carrying that exact line agrees. Where a full recording of the occasion exists, it is linked, so a passage can be read back in its own context rather than the channel's. A proposal that failed its check is still shown, marked as unconfirmed, because a lead is worth seeing as long as it is labelled as one.
The source library. Most of what the channel plays has already been published by the White House itself, in a video gallery of some 2,200 items running back to 20 January 2025. That gallery is indexed here, so a passage is first matched against it directly: where a run of uncommon word sequences from the feed occurs in one catalogued video and no other, the source is named outright and the link opens the original at the second the passage begins. That is evidence rather than inference, and it is what the matched in the White House video library tag means. Passages that match nothing in the library fall back to the search route described above, and stay clearly labelled as the weaker claim.
Why a transcript. A feed that replaces pool coverage is the administration's own edit of itself: what it repeats, whom it puts on camera, and what it leaves out are all measurable, but only if the audio becomes text. That is what this page is for.
Background: BBC · CNBC · the feed on YouTube