Internet Archive TV News Archive Data

It is the Internet Archive's broadcast-television collection: roughly 4.37 million US and international news broadcasts as of August 2026, each indexed with its closed-caption transcript so a phrase said on air retrieves the exact programme that said it. Coverage opens in March 2001 and runs continuously to the present - launch depth came from US national networks plus San Francisco and Washington DC locals, a donation of about 40,000 tapes from Marion Stokes's estate pulled in Philadelphia and Boston, and international broadcasters including M1 (Hungary), Press TV, Globovision, Syrian News and Zvezda now sit in the same index.

Within Datadory's fifteen-dataset broadcasting slice it holds a unique position: schedule catalogues describe programming, ratings currencies measure audiences, and this is the only product that preserves what news broadcasts actually said, searchable at caption level across a quarter-century. It scores 9 out of 10 on Datadory's quality rubric, with field definitions verified against sampled records spanning the collection's full range.

Get a sample of this dataset and see your channels, dates and phrases resolved against live rows.

One row per broadcast programme, exactly as it arrives:

identifier : CNN_20010314_050000_Larry_King_Live
title      : Larry King Live : CNN : March 14, 2001 5:00am-6:00am EST
date       : 2001-03-14

identifier : KQED_20260821_000000_BBC_News_The_Context_USA
title      : BBC News The Context USA : KQED : August 21, 2026 12:00am-12:30am PDT
date       : 2026-08-21

identifier : M1_20251130_180000_Idojaras-jelentes
title      : Időjárás-jelentés : M1 : November 30, 2025 7:00pm-7:05pm CET
date       : 2025-11-30
publicdate : 2025-11-30T18:00:00Z
description: Részletes időjárás-jelentés.

These are three real records spanning the shelf, and they make the structural argument better than any description could. A CNN hour from March 2001, a KQED airing of BBC News The Context USA from August 21, 2026 and a five-minute Hungarian weather report on M1 all land in one identical shape - which means a fact-checker verifying a 2001 claim, a media analyst tracking this month's cycle and an NLP team mining non-English captions query the same dictionary. Note the identifier doing quiet work: M1_20251130_180000_Idojaras-jelentes encodes channel, date and time before you read another field, so keys stay human-auditable even at four million rows.

Six verified fields carry every one of the roughly 4.37 million records, defined in the dictionary below. None of them is decorative. The identifier is a structured key rather than a random string, so channel and airtime survive even when titles change format; the pair of dates - broadcast versus ingested - is what lets you separate what aired from when it became retrievable.

Additional fields on request. The wider delivery carries more than the core row shape: caption text aligned to the video timeline for teams who need the transcript column beside the metadata, streaming clip references for embed workflows, borrow access flags on full broadcasts, and derived cuts such as per-channel or per-day counts. Each gets defined and exemplified against live records before your pipeline locks a schema, so the dictionary you buy is the dictionary you tested.

• Geo: Primarily United States national and local news, plus international broadcasters. Earliest holdings centre on San Francisco and Washington DC stations; the Marion Stokes donation added Philadelphia and Boston; recent results surface M1 (Hungary), Press TV, Globovision, Syrian News and Zvezda.
• Temporal: Records dated from March 2001 through the present, ingested continuously - a quarter-century of news cycles in one consistent frame, new footage landing days old beside material twenty-five years old.
• Granularity: One record per broadcast programme, typically half-hour to hour-long segments, searchable at caption-text level rather than only by title or channel.

That footprint is the pitch: schedule catalogues describe what was meant to air, ratings currencies count who watched, and this archive preserves what was actually said - searchable by the words in the captions.

API, files, or your warehouse. Daily, weekly, or hourly.

Pick the slice and set the cadence to match the decision being fed. Fact-checking desks want the newest broadcasts rotating in fast, because a claim dies inside a news cycle; historical researchers take the March 2001-onward backfill once and append as it grows; ML teams pin a fixed snapshot for reproducible training runs and refresh on their own clock. Because broadcast date and publicdate ride on every record, incremental pulls key cleanly without double-counting. The sample comes first, so channels, span and cadence are settled facts before any commitment.

• Journalists and academics run claim-checking as a lookup: quote in, moment-on-air out, with citable footage behind every hit. The journalists academics use cases page shows the workflow against the rest of the broadcasting shelf.
• Competitive intelligence and product teams watch when a rival, a category or a phrase entered the news cycle and how coverage curved afterward. The competitive intel product teams use cases page covers the patterns.
• Data scientists mine a quarter-century of aligned video-transcript pairs for media-measurement and NLP work - the data scientists use cases page has the model-side detail, and ml model training shows the broader pattern.
• Market researchers quantify coverage themes for client briefs with a primary-source corpus instead of a clipping service; citation-grade research is the general case.

Persona fit is lopsided by design: sales teams find no prospecting signal in broadcast captions, and e-commerce operators find no pricing or assortment content - the archive is unambiguous about what job it does.

Three notes worth having before you request a sample. First, this is the news slice of the Internet Archive, not the whole non-profit: the Wayback Machine's web captures and the Digital Library's scanned texts are separate collections with separate dictionaries, catalogued under their own industries. Second, full broadcasts travel under a borrowing model while short clips stream freely - if your workflow needs embeddable moments, scope the clip layer at sample time; if it needs whole programmes, say so, because the two arrive differently. Third, the station roster is not published as a list - it is derivable from the records themselves, which is precisely what a scoped sample settles: name the channels you assume you need and the sample confirms them against live rows.

Notes that pair well with this page:

• The broadcasting data hub collects all fifteen datasets in the industry, this one included.
• Schedule counterpart: TVmaze API - TV Show and Episode Database - episode and cast metadata for general programming; pair it when your subject is the programme guide rather than what was said.
• UK demand side: BARB Weekly Top 50 Shows and Viewing Data - what UK audiences actually watched, joinable to coverage for reach-versus-tone analysis.
• US dial registry: FCC TV Query System (LMS / CDBS TV Station Database) - facility-level reference data to enrich channel call signs.
• Head-to-head: TVmaze vs the Internet Archive TV News Archive - the side-by-side contrast between schedule metadata and caption-searchable news.
• Vocabulary: closed-caption search - how transcript-level retrieval differs from title-level search, anchored on this archive.