It’s Kelly, technical co-founder here at Signl.fm. I’m happy to announce that we’ve just released robotsroundrobin.com. It’s a game about refining speech-to-text.
Kim (our CEO, my cofounder) and I are both from video games. We started out at different times, since I’m a little baby and Kim is so old, but that knowledge has leaked into all of our work. We decided it was time to flex that side of our experience a little more than usual, and thus, Elliot was conceived.
The concept is simple: players are shown three automatically-generated transcripts, and asked to select the one they prefer. The vocabulary, “Which do you prefer?” is very important. This isn’t necessarily about correctness—we can check that with an algorithm. We want to know which transcripts are more preferable to average listeners. For example, I lean towards the robot that most accurately transcribes swearing. It makes me happy to know the robots understand vulgarity. We are excited to see what trends we uncover from all of you.
Under the hood, the robots represent different speech-to-text recipes. Sometimes, they’re not the most accurate when it comes to… converting speech into text. This is to be expected, the technology is young, and they haven’t had a super-diverse group of voices to train on. RobotsRoundRobin will help teach these machines to better understand podcasters, wherever their voices come from.
As I said earlier: this is more game-y than anything we’ve done publicly here at Signl.fm. I was inspired by the post-episode screens in Telltale’s games (The Walking Dead, Wolf Among Us), and the decision-making sections of Catherine. Both enforce to the player how their choices weigh up with the community, giving them a sense of camaraderie with their fellow players. We used this idea as a primary focus for the design.
All this data will make the robots smarter. Expect another blog post in a few weeks when I get to show off what we’ve learned.