Industries / Media, arts and recreation / Media and entertainment / Podcasting
Editing
Turning raw recordings into a finished episode: cutting, cleaning up the sound, mixing and setting loudness, and more and more using AI to make voices, translations or whole shows.
7
Milestones achieved, from 2006 to 2025.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2006
Podcast tools in the box
Apple's GarageBand 3 came in the iLife '06 suite, at a suggested US price of $79 for the whole suite, and Apple called it "a complete solution for creating professional quality Podcasts." It added radio-style sound effects, jingles, chapter artwork and web links, and could record guests interviewed over iChat AV straight into the editor. Free editing software already existed: Audacity, an audio editor that Dominic Mazzoni started in 1999 as a graduate student at Carnegie Mellon University, grew out of research with his adviser Roger Dannenberg.
Credit
Apple; Dominic Mazzoni and Roger Dannenberg (Audacity)
2015
A loudness target for streams
Loudness is measured in LUFS (loudness units relative to full scale, also written LKFS), one number for how loud a programme sounds on average to the human ear, with numbers closer to zero louder. In October 2015 the Audio Engineering Society's technical document AES TD1004 recommended a target of −16 to −20 LKFS for streamed audio and for files downloaded and played later, according to the ITU's 2018 report on internet loudness, which also found "no common target loudness level" among internet services. Apple asks podcasters for about −16 LKFS, give or take 1 dB, with peaks no higher than 1 decibel under the loudest level a digital file can hold, and presents these as recommendations, not requirements.
Credit
Audio Engineering Society; Apple
2017
Editing the transcript
In December 2017 Descript, spun out of Andrew Mason's audio-tour company Detour, raised $5 million from Andreessen Horowitz. Its editor laid out a recording's transcript like a word-processor document, and cutting a word from the text cut it from the sound, so editors could work on words instead of waveforms.
Credit
Andrew Mason and Descript
2019
Cloning your own voice
At its MAX conference in 2016 Adobe showed Project VoCo, which could type new words into a recording in the speaker's voice from about 20 minutes of their speech; Adobe said the demonstration raised "concern over how it might be misused" and that it "may never become a product feature." In September 2019 Descript bought Lyrebird, a Canadian voice-synthesis start-up, and launched Overdub, which fixes a line by typing it in a synthetic copy of the podcaster's voice. To block clones of other people, Overdub learns a voice only from sentences the user reads aloud from text generated on the spot, so nobody can upload existing recordings of someone else.
LessonEditing with a synthetic voice reached a shipping product once a consent check was built into how the voice is made.
Credit
Descript (Andrew Mason); Lyrebird, founded by three PhD students at Mila, the Montreal AI institute; Adobe (VoCo)
2023
Translation in the host's voice
On 25 September 2023 Spotify began a pilot that translates episodes into other languages in the host's own voice, using voice-generation technology OpenAI had recently released, starting with Spanish versions of episodes by Dax Shepard and Monica Padman, Lex Fridman, Bill Simmons and Steven Bartlett, with French and German to follow. In October 2023 PodcastOne began Spanish versions of its shows with the dubbing firm Rask AI, and Podnews noted that its own editor had been machine-translated into Spanish almost two years before Spotify's pilot. Spotify called its version a pilot, and this research found no later Spotify announcement that widened it.
Credit
Spotify and OpenAI; PodcastOne and Rask AI
2024
Shows made by machine
On 11 September 2024 Google added Audio Overview to NotebookLM, its AI notebook: one click turns uploaded documents into a downloadable conversation between two AI hosts, which Google said spoke only English and "sometimes introduce inaccuracies." In an independent test in the American Geophysical Union journal Perspectives of Earth and Space Scientists, Ian Flynn of the University of Pittsburgh and Sean Peters of Middlebury College turned three papers on Martian volcanoes into overviews; every overview contained errors, the errors came at the end, and one jumped from a volcanic feature to liquid water and possible life on Mars, which the paper never claimed.
LessonA machine-made episode still has to be checked against its sources, the end of the episode most of all.
Credit
Google (NotebookLM)
2025
The AI show factory retreats
In September 2025 Podnews, citing The Hollywood Reporter, reported that Inception Point AI, run by the former Wondery chief operating officer Jeanine Wright, was publishing more than 3,000 AI-made episodes a week at an apparent $1 each with four content staff, and that some shows did not say they were made by AI. Some of its Spanish episodes later contained leftover machine text such as "I'm sorry, I can't assist with that," and in December 2025 the Castro app hid the company's shows from its search and charts. In May 2026 the company, with about 15,000 shows, stopped publishing automatically; by September it was called Liquid Studios, its Quiet Please network had 300 active shows that it was no longer trying to earn money from, and Wright said some of the publicity had been "a PR strategy to get attention."
LessonPublishing without an editor's check put machine errors in front of listeners, and apps answered by filtering the shows out.
Credit
Inception Point AI (Jeanine Wright); reported by Podnews
Watch for
Raw tape to rough cut
Descript says its AI agent, Underlord, starts from "a recording and a plain request" and can "execute entire editing workflows" from a single prompt, with a rough cut as the example, removing filler words when asked; one creator Descript quotes says his prompt "does about 90% of the work." Descript also warns that the agent "can misunderstand or misinterpret what you tell it," and does not claim a finished cut in one pass.
Achieved whenan outside party runs it on raw interviews and publishes how its cuts compare with a human editor's, for example in a blind listener test.
Claimed by
Descript, in a guide updated July 2026
A podcaster with no human
At The Podcast Show in London in May 2026, Jeanine Wright announced what she called "the world's first totally autonomous podcaster." People would review its content for now, she said, but the company may "eventually" "largely remove the human in the loop of the review."
Achieved whenan outside party confirms that a show has published for months with no human review and reports its error rate.
Claimed by
Inception Point AI (now Liquid Studios), May 2026
Labels for AI-made audio
In May 2026 Spotify said AI "can't be used to impersonate others using voice clones" and began letting podcasts be marked "verified"; Podnews reported that Spotify still had no policy on shows made entirely by AI. In December 2025 Castro said it would reconsider hiding Inception Point's shows if their feeds clearly marked the content as AI-made.
Achieved whena major listening app shows listeners a label on AI-made episodes, based on the feed or the audio, and outside reporting confirms the label is applied.
Claimed by
Spotify, May 2026
Open
01
Blind test of AI edits
Collect raw interviews, either your own recordings made with the guests' consent or Creative Commons tape. Have an AI editor and a human editor each cut every interview to the same length, set both versions to the same loudness, let listeners hear each pair blind and pick one, and log every cut each version made.
The briefCredit · unclaimed02
A check for meaning-changing cuts
A tool that takes the raw recording and the finished episode, lines up the words in both, finds every cut and flags those that change what someone said: a dropped "not," an answer moved under a different question, two sentences joined into one claim. Test it on planted edits.
The briefCredit · unclaimed03
A loudness census of podcasts
Measure the integrated loudness, the average over a whole episode, and the true peak of a large random sample of recent episodes from many countries and languages, and publish the share that falls within Apple's −16 LUFS, give or take 1, and within the AES range of −16 to −20.
The briefCredit · unclaimed
Open 01 · the brief
Blind test of AI edits
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Collect raw interviews, either your own recordings made with the guests' consent or Creative Commons tape. Have an AI editor and a human editor each cut every interview to the same length, set both versions to the same loudness, let listeners hear each pair blind and pick one, and log every cut each version made.
Why it is still open
Descript's own guide to Underlord does not claim a finished cut from one request and warns that the agent can misread instructions. This research found no published blind comparison of an AI-edited and a human-edited podcast; the closest work was blind tests of AI voices against human voices in radio promotions and audiobooks, which test voices, not cuts.
Built with what exists
Descript Underlord and Riverside's editor for the AI cuts; WhisperX, an open speech-to-text tool that times each word, to list what each version removed; ffmpeg's loudnorm filter to set both versions to −16 LUFS; a simple web page that plays each pair in random order.
First step
Record or collect five raw interviews and recruit one human editor.
Done looks like
20 interviews and at least 100 listeners, with a published tally of preferences for each interview and the full list of cuts, marking any that changed meaning.
Open 02 · the brief
A check for meaning-changing cuts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that takes the raw recording and the finished episode, lines up the words in both, finds every cut and flags those that change what someone said: a dropped "not," an answer moved under a different question, two sentences joined into one claim. Test it on planted edits.
Why it is still open
Research on detecting edited speech, such as the AiEdit set of about 140 hours (2026), covers additions, deletions and changes made by AI speech-editing models and asks whether audio was altered; its description does not mention judging whether a cut changed the meaning or comparing against the original recording. This research found no tool that compares raw tape with an edited episode for meaning, and detectors of audio splices find joins but cannot say what a join did to the meaning.
Built with what exists
WhisperX, whisper.cpp (see the Note "A first stable number for local AI") or ElevenLabs Scribe v3 (see the Note "Fewer word errors") for word timings; pyannote, an open tool that works out who spoke when, so each question stays paired with its answer; Anthropic's Claude Fable 5.2 (see the Note "Three working devices") to judge each flagged cut against the original context.
First step
Take one Creative Commons interview, make 20 planted cuts by hand, and line up the words of both versions.
Done looks like
An open test set of at least 200 planted cuts that change meaning and 200 harmless ones (ums, repeats, tangents) on Creative Commons tape, with the tool's catch rate and false-alarm rate published.
Open 03 · the brief
A loudness census of podcasts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Measure the integrated loudness, the average over a whole episode, and the true peak of a large random sample of recent episodes from many countries and languages, and publish the share that falls within Apple's −16 LUFS, give or take 1, and within the AES range of −16 to −20.
Why it is still open
Apple's figures are recommendations it does not enforce, and the ITU found in 2018 that internet services had "no common target loudness level." This research found many guides that repeat the −16 target but no published measurement of real episodes at scale; the closest was a January 2013 blog post by Auphonic, which this research could not open.
Built with what exists
The Podcast Index API to draw the sample; ffmpeg's ebur128 filter or the pyloudnorm library, both based on the ITU-R BS.1770 standard. Publish numbers only, delete the audio after measuring, and keep download rates low out of respect for hosts' terms.
First step
Pull the 200 newest episodes from the Podcast Index and measure them.
Done looks like
At least 5,000 episodes measured, with a public table and script, and the share within 1 LU of −16 broken down by country and category.