Korva is a scroll of the internet’s best ideas.
Observed listening beats survey answers
Sensors inside the apps
The software is an SDK, a software development kit, that sits inside publisher apps where people read articles, listen to music, and play podcasts. It uses the phone's hardware sensors, the gyroscope and the accelerometer among them, to work out behavior: where someone is, what they're doing, and whether location is on or off. A single person can come out of it as a regular shopper and international traveler with a desk job.
Matching the records
Two more sources join that behavior data. One is the demographic profile a person gave the app when they logged in, their age and gender, which is the publisher's own first-party data. The other is the content itself: every podcast listened to, every stream, every article read.
Matching those records builds neighborhoods of content consumption across millions of users. The result is far larger than any survey, and it rests on what people really consumed.
Labeling the podcasts
Podcasts get labeled from the other side. Embedding models read the whole RSS feed, the large text file holding a podcast's description, keywords, and other core metadata, and map it to the most relevant contextual labels.
How do phone sensors reveal what someone is doing?
An accelerometer measures how a phone speeds up or slows down along three axes, and a gyroscope measures how it rotates. Walking, riding a bus, and sitting at a desk each produce a distinct pattern of those readings. Software trained on labeled examples of each activity can recognize the pattern without seeing a screen or hearing a sound. Phone operating systems ship their own activity-recognition features built on the same sensors.
Why can embedding models label podcasts without keywords?
An embedding model turns a block of text into a long list of numbers, arranged so that texts with similar meaning end up close together. A description about sourdough and one about bread baking land near each other even when they share few words. Each label can be turned into numbers the same way, and a podcast gets the labels whose numbers sit closest to its own. RSS itself is a feed format from the late 1990s that podcasting adopted in the early 2000s.
More from PodBiz | Where's The Money In Podcasting?
