clock menu more-arrow no yes mobile

Filed under:

Audible is using machine learning to let romance novel fans ‘skip to the good part’

New, 5 comments

Yes, the feature finds sex scenes, but it’s also about using machine learning to help break down a library for a voracious audience


My husband likes to tell a story from his college years: he was working in data entry, next to a woman with a copy of Anne Rice’s erotic novel The Claiming of Sleeping Beauty sitting on her desk. When a supervisor walked by and noticed it, she snapped it up, saying, “This is a romance novel? Hey, let’s look for one of the good parts!” Then she started flipping through the pages. Finding page after page of S&M scenarios and graphic sexual activity, she quickly put the book back down. “Oh,” she said, visibly embarrassed, “it’s all good parts.”

That doesn’t apply to most romance novels, though. Erotica designed around women’s tastes is more often about context, the slow build, and servicing a variety of fantasies besides the obvious sexual ones. Which has never stopped curious bystanders from pulling the “Let’s find the good parts!” move. It’s extremely common for romance novel readers to have experiences like this, where other people grab their books, hoping to sneak a little thrill by finding the smuttiest parts of whatever they’re reading. But the “good parts” of a romance novel are different for different readers. It’s a comfort genre, and people find their comfort in different places. The audiobook subscription service Audible figured that out ahead of the launch of their new Audible Romance service, and they’re using data mining and machine learning to figure out exactly what individual customers want, and to get them to it faster.


Audible’s chief content officer, Andy Gaies, tells The Verge that Audible Romance was plotted around the things that make romance novel fans different from other readers. “Romance fans are a unique group,” he says. “They’re some of the most savvy, dedicated, sophisticated, and voracious consumers of content within their genre. Many romance fans will consume four, five, six books a week.” So while the normal Audible service grants subscribers one audiobook a month, with the option to buy more at a discount, Audible Romance is an all-you-can-read service, available either as a standalone subscription or an add-on for Audible users.

But that means new users are starting with a deluge of content, which provably makes consumers impatient and pickier about any perceived barrier to their enjoyment. So Audible Romance comes with a couple of special features: a “Steaminess” score that lets subscribers filter books by their graphic content, and a “Take Me To The Good Part” feature that lets them jump to pre-selected scenes within certain criteria.

That sounds like Audible has bookmarked the sex scenes in its audiobooks. But the feature is much more nuanced, and developing it involved studying user habits and tastes. Audible Chief Financial Officer Cynthia Chu, who heads the company’s data science team, says Take Me To The Good Part highlights scenes in 10 categories: First Meeting, Flirty Banter, Sexual Tension, First Kiss, I Want You, Trouble in Paradise, It Might Be Love, Declaration of Love, and Proposal, as well as the sex scene category, Hot Hot Hot. She describes these categories as breaking down “moments in a typical women’s title where we know fans go back to relive them time and time again.”


Generally, a given book will only feature scenes pegged to five or six of these categories at most, but users can individually select which kind of scene they want to jump to in a given book. “The good parts are really up to the listener’s ears,” Chu says. “So we want them to choose what those good parts are. It’s different for everybody.”

“As we tested this, we found customers want to use it in very different ways,” Gaies says. “Some are using it for content discovery: ‘I want to listen to this part before I invest in this book.’ Some are using it as a consumption mechanism: ‘I just want to listen to a lot of moments around proposals, because that’s really exciting to me.’ And some are using it to revisit moments from their favorite listens: ‘I’ve listened to this book in the past, and I just want to go back and hear that flirty banter again, because it’s something I remember and loved, and it’s an easy way to get it back there.’”

Chu says her team used machine learning to isolate scenes in those 10 categories, creating an algorithm to scan for keywords associated with certain moments — “blush” for Flirty Banter, “kiss” and “embrace” for First Kiss, “ring” for Proposal — and then identify how they’re used in relation to each other. She calls those word relationships the “secret sauce” that makes the algorithm accurate.

“The main thing to remember is that the algorithm doesn’t just look at words by themselves,” Chu says. “It looks at context. ‘Making’ and ‘love’ are two words — if you look for them by themselves, your results may not make any sense. The algorithm is looking at groups of words and making sure we’re finding the right contexts.”

Once the computerized scan categorizes scenes, they’re handed to human editors for vetting, to prevent misidentification and to give feedback to the data science team on how well the searches are functioning. But Gaies says the machine learning system has largely been accurate, to a “really, really impressive” degree. He says the human curatorial process for Audible Romance is mostly about “how we talk to customers,” which includes breaking down books in a wide variety of categories, including setting, style, and character types. Professions like “hockey player” get their own Audible Romance pages, packed with titles like Pucked Up, Pucked Over, and Hot As Puck. There’s a noticeable sense of fun in the categories and breakdown descriptions, the kind of impish humor that computers can’t mimic yet — like the Character Types launch page description that invites users to “obsess over lovers from vampires to viscounts to Vikings.”


The sorting categories for themes and professions are human-curated, but Gaies says the Steaminess scores aren’t — “Those are algorithm-direct to the consumer.” So a computer program is deciding which of five categories a given romance novel falls into: Sweet, Simmering, Sizzling, Hot Damn, or O-O-OMG. That’s one area where Audible’s staff is looking forward to user feedback, so they can see how well the categorization worked, and whether listeners are getting the appropriate level of steaminess. “We’re likely going to get a few of them wrong,” Gaies says. “Figuring out how we can use that feedback loop to make the Steaminess score more accurate is really important early on.”

Ultimately, the goal is to help users sort through Audible’s library faster and find the books that most fit their tastes. “When you have a broad selection of content in an access-based service, it’s always difficult to help someone find that next listen,” Gaies says. “We had this perfect mix of technology and data science, along with some really bespoke editorial love and curation. I think this package solves those barriers we’ve identified, to why customers haven’t embraced the genre in greater numbers in audio.”

Romance novels may be uniquely suited to this data analysis approach, because they can be somewhat formulaic. Like many other genres, they have their patterns and their reader expectations. Rice has called her Sleeping Beauty books an experiment in “designing a book where you didn’t have to mark the hot pages, where every page would be hot.” But most romance novels pay more attention to the stages of relationships, as codified in decades of American entertainment in a wide variety of media.

So there’s certainly a question of whether Audible could expand the data-driven “find the good parts” approach to its other genres, letting listeners preview fight scenes in thrillers, or gory scenes in horror novels. Even if some romance readers do want to judge a book by jumping straight to the proposal scene and seeing how it fits their fantasies, that’s no guarantee that, for instance, mystery readers want a feature that will let them hop to the moment where the murderer is revealed. As far as plans to expand the machine categorization feature to the rest of Audible, Chu is positive, but noncommittal. “We’re an innovative company,” she says. “Every day, we use these algorithms we’ve developed internally, and look for use cases where we can enhance the customer content experience.”

For the moment, they’re content to have these features as a selling point for the new service, and a way to show off their capacity for giving users exactly what they most want. “It’s a way to humanize the vast data we have in all that text,” Chu says. “It just comes through a lens that’s respectful to the listener.”

Audible Romance launches on November 1st. According to Gaies, the initial launch catalog will include “over 10,000 titles,” and Take Me To The Good Part will initially be available for 100 of those books.