Assertion Supported AI assessment confidence: 90% certainty 3/5 debate potential 1/5

Malhotra: Earliest AI Video Synthesis Models Needed Over 20 Hours of Training Data

Ashray Malhotra · EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra · Jul 23, 2025 · at 18:22

Ashray Malhotra, co-founder of generative video company Rephrase.ai, explains why early synthetic media research relied on public domain video speeches of Barack Obama.

0:00 / 0:37exact quote · 37.2s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“The earliest models needed a lot of training data to be able to train. And Obama used to give a presidential speech every week. Which was put in public domain by the white house. So, and he gave that, I think for four to five years. So he basically had say, you know, a few minutes of his video every, every, every week for like a very long time. So that added up to, I think like, I may be getting the number wrong, but over 20 hours of data which is what the earliest models needed to be able to create a model that can synthesize space and voice of a person.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Early video synthesis models required over 20 hours of training data to synthesize a person's face and voice, which Rephrase.ai sourced from Barack Obama's weekly public domain presidential speeches.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

Published research literature confirms that seminal early deep-learning talking-head and audio-visual synthesis models relied on public-domain footage of Barack Obama's weekly presidential addresses. Specifically, the University of Washington's 2017 benchmark paper 'Synthesizing Obama' utilized 14 to 17 hours of weekly address footage, while MILA's 2017–2018 'ObamaNet' model used 17 hours of the weekly speeches to synthesize both voice and facial movements. Subsequent literature, including work published at AAAI 2022, compiled up to 20 hours of this public-domain dataset to train audio-visual synthesis models, aligning directly with the speaker's account and the resolution criteria.

University of Washington's 'Synthesizing Obama' (July 2017) utilized 14 hours of filtered footage (derived from ~17 hours across roughly 300 weekly presidential addresses) placed into the public domain by the White House.
MILA's 'ObamaNet' (December 2017 / January 2018) trained both text-to-speech voice synthesis and facial lip-sync using 17 hours of video from Obama's weekly addresses.
Follow-up academic literature, including a 2022 AAAI publication on single-speaker audio-visual correlation learning, specifically documented using 20 hours of Obama's weekly presidential address footage.
Under 17 U.S.C. § 105, official White House weekly presidential addresses are works of the U.S. government and enter the public domain upon publication.

Checked against

[1] medium.com [2] eurekalert.org [3] researchgate.net [4] engadget.com [5] eurekalert.org [6] futurity.org [7] washington.edu [8] medium.com

Synthesizing Obama paper hours of training data weekly presidential address "Synthesizing Obama: Learning Lip Sync from Audio" "hours" "weekly" Synthesizing Obama Learning Lip Sync from Audio hours video weekly addresses Synthesizing Obama 14 hours 17 hours 20 hours video "Synthesizing Obama" "17 hours" "300" "weekly" "Synthesizing Obama: Learning Lip Sync from Audio" "14 hours" OR "17 hours" OR "hours of" "Synthesizing Obama" "14 hours" OR "17 hours" OR "hours of" "weekly address" filetype:pdf

assessed as of 2026-09-08

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

Barack Obama rephrase.ai White House

⚑ Spot an error? Report it

More from Ashray Malhotra

Prediction Not checkable as stated

Malhotra: A significant number of future pixels will be generated, not recorded

“I believe most, okay, the most, but a significant number of pixels in the world in the future are going to be generated. And not actually recorded.”

Ashray Malhotra Jul 23, 2025 ▶ 46:25 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Opinion

Malhotra: Visual AI Models Still Await Their Revenue-Driven ChatGPT Moment

“I believe that the, you know, the chat GPT moment for the image, video, and three models is still to happen. You could make a claim that an image that is kind of happened, but no one's making billions of dollars in revenue on image models. The way people are m…”

Ashray Malhotra Jul 23, 2025 ▶ 51:21 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Assertion Not checkable as stated

Malhotra: Rephrase.ai's Synthetic Media Campaigns Never Failed, Unlike Competitors

“None of our campaigns went wrong, and I know many of our competitors have had that.”

Ashray Malhotra Jul 23, 2025 ▶ 20:10 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Assertion Supported

Malhotra: Cadbury's AI Campaign Won India's First Cannes Titanium Lion

“I think India won its first ever cons like major cons awards in the, which are the equivalent of Oscars. Yeah. We won Titan lion.”

Ashray Malhotra Jul 23, 2025 ▶ 37:05 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Insight

Malhotra: AI base technology is currently far ahead of the application layer

“It is a rare moment in history where your technology is so far ahead of the application layer that there is so much that we need to do here to catch up.”

Ashray Malhotra Jul 23, 2025 ▶ 50:42 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Assertion Not checkable as stated

Malhotra: Rephrase.ai Never Cloned Anyone Without Permission Beyond Early Demos

“And outside of a few demos like that, which we did in a very, very early days of our journey, we almost, like, as soon as the technology started to get good, we never cloned anyone without their permission.”

Ashray Malhotra Jul 23, 2025 ▶ 19:56 EP12: Synthetic Yet Real - The New Age of Personalized Video Content ft Ashray Malhotra

Malhotra: Earliest AI Video Synthesis Models Needed Over 20 Hours of Training Data

More from Ashray Malhotra

Turn any episode into a week of clips.