Industries / Media, arts and recreation / Media and entertainment / Radio broadcasting
Studios
Studio work is making the sound and keeping it: microphones, recorders and editing, the systems that carry audio around the building and play programs out to the transmitter, and the archives of what was aired.
4
Milestones achieved, from 1947 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1947
Shows go on tape
Bing Crosby wanted to stop doing his weekly show live. NBC refused, and ABC agreed to let him pre-record if the quality matched a live broadcast, but the transcription discs then used for radio sounded worse and were hard to edit. The engineer John (Jack) Mullin had brought two German AEG Magnetophon tape recorders back from the war and improved them. In 1947 he recorded the first show of the 1947–48 season of Crosby's Philco program on tape, and Crosby sent Ampex a $50,000 check so it could build machines and invested in 3M for a supply of tape. Tape let producers cut out mistakes, and other radio programs followed, by the IEEE's history.
Credit
Bing Crosby; John T. Mullin; Ampex (Alexander Poniatoff); 3M
2013
Audio networks interoperate
Studios had begun sending audio between rooms over computer networks, using several competing systems. In September 2013 the Audio Engineering Society published AES67, a standard that picks a subset of existing technologies that every system can share, so that equipment from different makers can exchange professional-quality audio (16 bit, 44.1 kHz and higher) with under 10 milliseconds of delay. The AES began the project in December 2010 and worked with the European Broadcasting Union from 2012.
Credit
Audio Engineering Society task group SC-02-12-H, led by Kevin Gross; European Broadcasting Union
2020
Hosts broadcast from home
When the coronavirus pandemic closed workplaces in spring 2020, many stations kept going with hosts in makeshift studios at home and only a few staff in the building. WTTW reported in April 2020 that Omar Ramos of Latino Mix 93.5 and many other Chicago DJs were working this way.
Credit
station hosts and engineers, among them Omar Ramos (Latino Mix 93.5) and Lin Brehmer (WXRT)
2025
A cloned host on air
CADA, a Sydney station owned by ARN, ran Workdays with Thy for four hours a day on weekdays for about six months. Its host was AI-generated, made with ElevenLabs voice software, and her voice and likeness were modelled on an employee in ARN's finance department; listeners were not told. It came out in April 2025 after the Sydney writer Stephanie Coombes asked who Thy was, and ARN told the Australian Financial Review that it had been "trialing AI audio tools," Global News reported.
LessonA cloned voice can host a show for months without listeners noticing, and the objections came when they learned they had not been told.
Credit
ARN (Australian Radio Network); Stephanie Coombes, who asked the question
Watch for
Transcripts for unlabeled tapes
The American Archive of Public Broadcasting (AAPB), run by the Library of Congress and GBH, holds more than 271,000 digitized items and has 2.0 full-time staff for cataloging. It says it now uses Whisper speech recognition to transcribe recordings that came with no description, among them about 2,000 of the 8,000 radio records from WIPR, Puerto Rico's public broadcaster, so that catalogers have something to read when they describe them. It also says the transcripts contained hallucinations (text the model made up), especially over instrumental music.
Achieved whenthe AAPB or an outside evaluator publishes how many records were completed this way and how accurate they are.
Claimed by
AAPB, April 2026
Open
01
Draft catalog records for radio
A tool that listens to uncatalogued radio recordings and drafts item-level catalog records (program title, date, people, places and subjects), checks names against outside sources, and shows an archivist the evidence for each field.
The briefCredit · unclaimed02
Score live radio captions
A published test of automatic live captions for radio, measuring accuracy, delay and readability on real live programs (news, call-in shows, sports), with ratings from deaf and hard-of-hearing listeners.
The briefCredit · unclaimed
Open 01 · the brief
Draft catalog records for radio
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that listens to uncatalogued radio recordings and drafts item-level catalog records (program title, date, people, places and subjects), checks names against outside sources, and shows an archivist the evidence for each field.
Why it is still open
At the AAPB, Whisper transcripts help catalogers, but the catalogers still write the records, and the archive warns that listening alone "cannot tell you the proper spelling of a proper name." The Library of Congress's Radio Preservation Task Force has built a database of more than 2,500 radio collections in archives across the US. This research found no published test of AI-drafted radio catalog records against records written by archivists.
Built with what exists
Speech-to-text such as Whisper or ElevenLabs Scribe v3 (see the Note "Fewer word errors"); an AI model with web search, such as Claude Fable 5.2 (see the Note "Three working devices"), to propose titles and check names; the CLAMS tools from Brandeis University, which the AAPB already uses; the AAPB's public catalog for test items.
First step
Take 30 catalogued radio items from the AAPB's public website, hide their records, draft new ones, and score them field by field.
Done looks like
On 300 radio recordings that archivists have already catalogued, with their records hidden, the drafts get the title, date and main speakers right for at least 80%, archivists accept at least 60% with minor edits, and the time per record is logged against cataloging from scratch.
Open 02 · the brief
Score live radio captions
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A published test of automatic live captions for radio, measuring accuracy, delay and readability on real live programs (news, call-in shows, sports), with ratings from deaf and hard-of-hearing listeners.
Why it is still open
NPR Labs, Towson University and Harris showed captioned radio on election night 2008 using stenographers at WGBH. In 2022 Current reported that WAMU in Washington was testing Enco's automatic enCaption captions on a web page and that KQED was pointing listeners to Chrome's Live Caption; KQED's Ernesto Aguilar named accuracy across voices and accents as a problem. GBH's Donna Danielewski said mainstream use would still need radios with caption displays and standards for sending captions in the radio signal. This research found no published accuracy or delay figures for live radio captions.
Built with what exists
Live station streams; streaming speech-to-text such as ElevenLabs Scribe or Whisper; reference transcripts typed by people; the word error rate method explained in the Note "Fewer word errors"; a simple caption web page for the listener panel.
First step
Caption one hour of a live news program from a public stream with two systems and score both against a hand-typed transcript.
Done looks like
30 hours of live programming in three formats captioned by at least two systems; word error rate and median delay published per format and per system; at least 20 deaf or hard-of-hearing listeners rate readability.