Beyond the subtitles: AI-Media’s AI-native transformation

AI-Media co-founder and CEO Anthony Abrahams explains how a Foxtel captioning start-up shut down its last human captioner and rebuilt itself around AI, and why owning the encoder at the edge of the broadcast signal made it possible.

From a Foxtel Tender to a Global Business

AI-Media was founded in 2003 by Tony and his co-founder Leonie Jackson, who was deaf and later left to become the first deaf CEO of the Deaf Society. Foxtel was the foundation customer. It had promised the deaf community it would introduce captioning when the service went digital, and it first hired the pair to consult on how to set up a captioning bureau.

Leonie’s argument was that the market was far bigger than the sceptics assumed. One in six Australians has a hearing impairment, and one in four speaks a language other than English at home, and captions help both groups. Foxtel, then run by Kim Williams, liked the idea but did not want another cost centre. He told them to set up a business and tender for the work.

Foxtel tendered the work in eight separate rounds over six months, and the young company won seven. Tony says the trick was understanding what pay TV needed, which was to fill a schedule as cheaply as possible. Caption files for shows that had already aired in the US or UK were sitting on servers and retailing for around $300 an hour. AI-Media offered $100.

Three Weeks From Running Out of Money, Three Times

Tony says he loves listed life, but he is blunt about what cash flow does to a founder.

The company was three weeks from running out of money on three separate occasions. The next pivots were into broadcast, winning the Channel 9 contract in 2014 and Channel 7 in 2016. In 2017 it became the launch partner for live captions on Facebook, which gave it a ticket into America. That was followed by a couple of acquisitions.

The contrast in fundraising is stark. Raising capital in 2014 took Tony 18 months and was, in his words, awful. Ten months after the IPO, AI-Media raised the money to buy EEG in 36 hours. Tony says the company would not have the business it has today if it had not been listed.

The Encoder at the Edge: Why EEG Was the Real Prize

AI-Media works across broadcast, government and enterprise, and broadcast, mostly in North America, is by far the biggest. In that world, everything starts at the encoder, the device that sits at the final point of broadcast, after the newsroom production, audio mixing and graphics are done, right before the signal goes to air. EEG has made captioning encoders since 1980, and when AI-Media bought it in 2021 it held about 80% of the US market.

The market’s reaction was that AI-Media was buying a low-growth hardware business that was already tapped out. Most of EEG’s business was hardware, but Tony saw something else.

Because it’s not about the encoders. It’s actually about the encoder being in the best place to generate the best quality live captioning.

— Tony Abrahams

Tony recalls that EEG had already built iCap, an IP network that let human captioners connect to encoders over the internet instead of dialling in by modem. It was free, a textbook strategy of making itself the default standard, and it helped EEG capture the market. In 2016 the founder’s son, Bill, who is now AI-Media’s chief product officer, put automated captioning on the system for the first time.

By 2019 Tony’s team knew that automatic captioning was coming and that the encoder was going to win, because it sits behind the customer’s firewall with access to everything. It is also very sticky.

No one unplugs an encoder once it’s in the rack, because the entire broadcast signal goes through it, and the entire broadcast signal will go to black.

— Tony Abrahams

At the time of the deal, LEXI was already a small but fast-growing part of EEG’s business, and it has grown sharply since.

Inside LEXI: Automation That Knows What’s on Air

In the past, live captioning at the encoder needed a human re-speaker who had trained the speech recognition. Tony says the modern version is a far more sophisticated AI analysis that looks not just at what is said but also at body language. That has led to a new product, LEXI AD, an audio description service for blind and low-vision viewers. He jokes that fund managers might want it to run sentiment analysis on CEOs at reporting season.

The LEXI orchestration layer also reads the scheduling and metadata sitting behind a broadcast: which story is next, when the commercial break starts and what is being promoted after it. Tony gives the example of Channel 7, an exclusive customer for ten years, which has more than 100 regional variants. LEXI reads the daily schedules so that a suburb name spelt one way in Brisbane and another in Melbourne comes up correctly on screen. For Fox in the US, AI-Media has just plugged into the automation system so captions stop when a Coke commercial with its own baked-in captions comes on, ahead of moving more Fox stations off human captioners.

LEXI Voice: Live Dubbing in a Day

LEXI Voice live-dubs broadcast content into other languages. Tony’s example is a Taiwanese news channel that wanted to hear in Mandarin, immediately, what Donald Trump was saying.

It was an inbound from them. And we had them up and running in a day.

— Tony Abrahams

AI-Media also ran Spanish voice tests during the World Cup, with Fox holding the US rights. These were tests rather than broadcasts to air. To make a Spanish dub work for live sport, the company delays the video feed so picture and dubbed audio stay in sync, because nobody wants video that turns up half a second behind the sound.

Simon jokes about how Netflix dubs bother him, and Tony says the next project, with NVIDIA and AWS, is live lip sync. AI-Media also works with AudioShake, which can strip out noise like drones from Olympic coverage and isolate individual speakers. That lets a commentator be pulled out, converted into Spanish and layered back over the original crowd noise, which Tony says is exactly what sports broadcasters want.

Switching Off the Last Human Captioner

AI-Media took the business to IPO entirely re-speaker based. It shut the gates on the last re-speaker on 31 December 2025, which cycled out that entire part of the business.

Alongside the shift, headcount has come down from a peak of 963 to about 150 today, from about 175 a year ago, and Tony expects it to settle around 125.

The hardest day of his time as CEO was tied to human error. In January 2022, Channel 7 called to say behind-the-scenes footage of news readers had leaked online. A captioner had sent something they thought was funny to three friends on WhatsApp at 11pm, and by 8am it had been viewed 800,000 times. Tony’s point is that the new LEXI encoders have removed the humans altogether.

New Hardware, a Security Rewrite and What Comes Next

Two new encoder form factors, the LEXI Text Encoder and the LEXI Voice Encoder, were launched at the NAB trade show in April. Tony says the point was to get them into a few customers’ hands rather than start hard selling. The feedback, including a request for a bigger screen, means the version shipping next month looks quite different. Customers replace hardware roughly every five years, so nobody is banging down the door for one tomorrow.

The biggest part of the investment in the new suite was a security rebuild. The software was rewritten from scratch, with the help of Claude Code and other AI tools. Tony warns that AI-powered attacks are about to get much worse, and describes a system that both blocks intrusions and constantly scans inside the network for outlier behaviour, like a vaccine. AI-Media throws AI cyber attacks at it to test it, and so far, he says, the new system has not failed. That matters for customers like a small regional news station with two encoders and one engineer. The build phase is now done, and as Tony puts it, they have built it and now have to sell it.

The Moat Is the Workflow, Not the Model

If speech-to-text is close to free, why would a customer like the NBA still pay? Tony says it is because the job is no longer about speech-to-text.

It’s actually about looking at AI across all elements of audio and video, reusing that information across multiple applications.

— Tony Abrahams

The suite covers text, voice, translation, recorded content, audio description and LEXI Insights, which captures what is broadcast so it can be queried like an LLM (“what happened on Channel 9 today?”). On the models underneath, Tony refines the label he is given.

I wouldn’t necessarily say we’re model agnostic, which would imply we don’t care which model we use. We do care. We rank them.

— Tony Abrahams

Some best-in-class voices can cost 20 times as much as Google’s, which Tony calls fine, so AI-Media does not compete in that layer. It coordinates the information, chooses which engine to send it to and strips out the noise around it.

Tony also points to a gap in the market. Education was once a very large part of AI-Media’s business in the re-speaker era and is now a small fraction of it. The answer is LEXI Access, where a student scans a QR code and picks captions in English or live Mandarin text or audio, piggybacking on the audio-capture equipment that universities already have.

The 12-Month Scorecard

Asked what evidence he would want to be held to in twelve months, Tony gave three things, without putting numbers on them:

• LEXI Text keeps growing.
• Encoder replacement picks up pace, with more old units swapped out next month than last month.
• LEXI Direct is live on three platforms. It puts LEXI inside another company’s environment, and launches in Amsterdam in two weeks with Grass Valley, making AI-Media the first language layer on its AMPP system.