The boom in popularity for podcasting has given a new voice to the world of spoken word content that had been largely left for dead with the decline of broadcast radio. Now riding the wave of that growth, a startup called Descript that’s building tools to make the art of creating podcasts — or any other content that involves working with audio — a little easier with audio transcription and editing tools, has a trio of news announcements: funding, an acquisition and the launch of a new tool that brings some of the magic of natural language processing and AI to the medium by letting people create audio of their own voices based on text that they type.
Descript, the latest startup from Groupon founder Andrew Mason, created as a spin-off of his audio-guide business Detour (which got acquired by Bose last year), is today announcing $15 million in funding, a Series A for expanding the business (including hiring more people) that’s coming from Andreessen Horowitz (it also funded the startup’s seed round in 2017) and Redpoint.
Along with that, the company has acquired a small Canadian startup, Lyrebird — which had, like Descript, also built audio editing tools. Together, the two are rolling out a new feature for Descript called Overdub: people will now be able to create “templates” of their voices that they can in turn use to create audio based on words that they type, part of a bigger production suite that also will let users edit multiple voices on multiple tracks. The audio can be standalone, or the audio track for a video.
(The video transcription works a little differently: When you add words, or take them out, the video makes jumps to account for the changes in timing.)
Overdub is the latest addition to a product that lets users create instant transcriptions of audio text that can then be cut and potentially augmented with music from other audio using drag-and-drop tools that take away the need for podcasters to learn sound engineering and editing software. The non-technical emphasis of the product has given Descript a following among podcasters and others that use transcription software as part of their audio production suites. The product is priced in a freemium format: no charge for up to four hours of voice content, and $10 per month after that.
In the age of market-defining, election-winning fake news aided and abetted by technology, you’d be forgiven for wondering if Overdub might not be a highway to Deep Fake City, where you could use the technology to create any manner of “statements” by famous voices.
Mason tells me the company has built a way to keep that from being able to happen.
The demo on the company’s home page is created with a special proprietary voice just for illustrative purposes, but to actually activate the editing and augmenting feature for a piece of their own audio, users have to first record a number of statements that are repeated back, based on text created on the fly and in real time. These audio clips are then used to shape your digital voice profile.
This means that you can’t, for example, feed audio of Donald Trump into the system to create a version of the president saying that he is awfully sorry for suggesting that building walls between the U.S. and Mexico was a good idea, and that this would not, in fact, make America Great Again. (Too bad.)
But if you subscribe to the idea that tech advances in NLP and AI overall are something of a Pandora’s box, the cat’s already out of the bag, and even if Descript doesn’t allow for it, someone else will likely hack this kind of technology for more nefarious ends. The answer, Mason says, is to keep talking about this and making sure people understand the potentials and pitfalls.
“People already have created the ability to make deep fakes,” Mason said. “We should expect that not everybody is going to follow the same constraints that we have followed. But part of our role is to create awareness of the possibilities. Your voice is your identity, and you need to own that voice. It’s an issue of privacy, basically.”
The developments underscore the new opportunity that has opened up in tapping some of the developments in artificial intelligence to address what is a growing market. On one hand, it’s a big market: Based just on ad revenues alone, podcasting is expected to bring in some $679 million this year, and $1 billion by 2021, according to the IAB — one reason why companies like Spotify and Apple are betting big on it as a complement to their music streaming businesses.
On the other, the area of production tools for podcasters is a very crowded market, with a number of startups and others putting out a lot of tools that all work quite well in identifying what people are saying and transcribing it accurately.
On the front of transcription and the area where Descript is working, rivals include the likes of Trint, Wreally and Otter, among many others. Decript itself doesn’t even create its basic NLP software; it uses Google’s, as basic NLP is now an area that has essentially become “commoditized,” said Mason in an interview.
That makes creating new features, tapping into AI and other advances, all the more essential, as we look to see if one tool emerges as a clear leader in this particular area of SaaS.
“In live multiuser collaboration, there is still no other tool out there that has done what we have done with large uncompressed audio files. That is no small feat, and it has taken time to get it right,” said Mason. “I have seen this transition manifest from documents to spreadsheets to product design. No one would have thought of something like product design to be huge space but just by taking these tools for collaboration and successfully porting them to the cloud, companies like Figma have emerged. And that’s how we got involved here.”