We spent the last two months getting more training data and making adjustments to improve the Qbrio AI’s performance. The new model is now live on the site, try it out!
Here is one success with the updated model. The show Suits played Desi Valentine’s very moving “Fate Don’t Know You” at the end of an episode in its sixth season. The Suitsteam used an especially stirring moment (3:14-3:15 in the song below) to create an emotional climax exactly when Harvey and Donna hold hands at the end of the episode, after which the picture cuts to black. I remember getting the chills when I watched that scene live and immediately looking up the music.
We were curious to see if the part of the song the Suits team chose – a good example of Grief pattern with the falling, wailing vocals and supporting Resolution pattern with the completed cadence – would also be the part of the song that Qbrio liked the most. Well, the highest Moment Score (0.87, arrow below) flagged by the AI was at 3:14, precisely the part the Suits team used in the climax of the scene. This seems like good validation to us that Qbrio is improving
At the same time, an interesting new problem has emerged with latest model update: the Qbrio AI likes applause a little too much. We caught this problem when we were testing Coldplay’s “Fix You”, which is one of the most popular submissions to online frisson playlists. Our team of listeners flagged the Proximity pattern into 2:54 in the video below as a moment that was especially reliable for frisson.
When we ran “Fix You” through our new model, however, the highest Moment Score in Qbrio’s analysis (0.91, red arrow below) was the crowd applauding at 4:38 in the video. The climax at 2:54 only got a Moment Score of 0.84 (green arrow below).
See what you think, but at least for me 4:38 is not the most chills-inducing part of the video. Its also not that helpful or interesting for the AI to recommend that artists simply put more crowd applause into their music. This is a tricky problem to fix, however, because we did not give the AI a lot of training data where applause was tagged as frisson-inducing. We think Qbrio likes crowd applause/whooping because applause has high tonal volume, through which it can create a Surround pattern (in fact if you listen carefully Coldplay’s team faded in applause noise in the 2 seconds before the climax at 2:54 too, enhancing the effectiveness of the crescendo there). So the AI is not entirely wrong, crowd applause can give listeners chills, but we think Qbrio likes applause a little too much at the moment.
We’ve got some more fine-tuning to do in the next model update.