Lyrebird

» see lyrebird.ai as it lived, on the wayback machine

~ the goodbye ~

In April 2017 a page went up with audio of Donald Trump and Barack Obama talking about a company called Lyrebird. Neither man had said a word of it. The site took a hundred thousand visits on its first day, and the title on the page that week was not coy about what was being sold: Lyrebird, an API to copy the voice of anyone.

Behind it were Alexandre de Brebisson and Jose Sotelo, working with Kundan Kumar, all doctoral students at MILA, the machine learning lab at the Universite de Montreal that Yoshua Bengio directs and that advised the project. The claim was that a minute of recorded speech was enough. Feed it sixty seconds of somebody talking and the model would extract what the founders called that person's speech DNA, then say anything you typed in their voice, with the intonation and the accent carried over and the emotion adjustable by hand.

For most of 2017 and 2018 this was the thing journalists wrote about when they wanted to frighten readers about audio. NPR ran it twice. The CBC ran it. Every piece worked the same way, which was to play the Obama clip and then wait. The word deepfake was barely in circulation yet, and the video version of the problem was still a year from the front pages, so for a while Lyrebird was the whole public argument about synthetic media by itself.

On 18 September 2019 Descript said it had acquired the company, alongside a $15 million round led by Andreessen Horowitz and Redpoint. Descript made podcast editing software and was run by Andrew Mason, who had previously founded Groupon. The Lyrebird team became Descript's research group and the standalone product was retired, with the voice cloning surviving as a feature called Overdub.

~ the chronicle ~

~ the ethics page ~

The thing that made Lyrebird unusual was that it argued with itself in public, on a page in the top navigation labeled Ethics, next to Careers and Contact. It opened with "With great innovation comes great responsibility" and then made an argument that has aged into something you still hear every few months from much larger companies.

The reasoning went like this. The technology was early but would improve fast and become widespread, and that was inevitable, so the only real question was how to introduce it. Their answer was to raise public awareness deliberately, which is what the Trump and Obama samples were for. The page said outright that those clips existed to warn people rather than to sell them anything. The second principle was consent: "We are stewards of your voice, but you control its usage: no one can use it without your explicit consent."

Then the part that reads strangest now. "Imagine that we had decided not to release this technology at all. Others would develop it and who knows if their intentions would be as sincere as ours." A company with three graduate students in it published that about a capability which, seven years on, costs nothing and runs on anybody's laptop.

~ where the domain points now ~

Lyrebird's site stopped serving its own content in the middle of 2019, with the final capture of the real homepage dated 19 June. By then the tagline had been sanded down from an API to copy the voice of anyone into "Ultra-Realistic Voice Cloning and Text-to-Speech", which is the same product described by a company that had hired a marketing person.

After the acquisition the domain became a redirect, and the redirect target is worth following. In November 2020 lyrebird.ai sent you to descript.com/overdub?lyrebird=true, the dead brand pointed at the feature that replaced it with a query parameter naming the corpse. By 2026 the target had changed again, to descript.com/research, which is where the Montreal team now sits. The brand survives as a signpost aimed at its own authors.

~ what descript kept ~

Mason was explicit about the constraint when Overdub shipped: you could only use it on your own voice, and he said the feature existed to save the tedium of re-recording and splicing after an editorial change rather than to make fakes. That is a real answer to the problem Lyrebird's ethics page had raised, and it only works because the capability sits inside an editor you pay for, attached to an account.

I did not follow this at the time. Audio synthesis was not my corner of the field and Montreal is a long way from Ankara, so what I have is the same thing anyone else has, which is the archive. The useful thing in it is the gap between two sentences the same company wrote about itself eighteen months apart. First an API to copy the voice of anyone, published to make sure you knew. Then ultra-realistic text to speech, published to sell you a seat.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »