A never-ending worry for our team is bias. We are constantly asking ourselves questions like: Do we have too much “Western” pop music in the dataset? Do we have too many submissions from male vs. female listeners? Are our annotators missing frisson moments because they hate certain artists? Is the AI learning overly “intellectual” patterns that require deep genre familiarity to appreciate (e.g. unusual nuanced cadences that only experienced listeners might get frisson from)? These are challenges that are never going to be fully solved, but that can be detected and monitored with data.
One heartening aspect of our data collection process is the genre diversity it has yielded. While it’s a truism among frisson researchers that the effect is genre-agnostic, its still striking to see this reality emerge in the crowdsourced data. We have seen everything from classical to country, salsa to screamo, gospel to grunge, opera to overtone singing, dream pop to death metal, folk to funk, post-rock to progressive house. I have also been impressed with the amount of international music submitted, and listening to the international music has been one of the most enjoyable parts of the project for me. It’s quite moving to get chills from a musical tradition you have never encountered before. I remember early in our project getting goosebumps from a piece of Tuvaan throat singing, where I had absolutely no idea what the lyrics meant and was disoriented by the new sounds, but nevertheless I still got chills (cue the clichés about music being a common language).
Here is a (small) sample of international submissions to our dataset:
We also love submissions from diverse genres because the more varied the data we feed to the AI, the more precise and robust the features it learns. It’s a much stronger result when Qbrio is shown a pattern manifesting itself in multiple genres. The AI can then figure out which genre-specific attributes to de-prioritize and better hone in on the genre-agnostic patterns producing frisson.
To end on a more speculative note, artists can even start purposefully creating new sounds and samples specifically designed to induce frisson. Researchers speculate that many of the frisson-inducing patterns in music today came about from humans mimicking sounds they encountered in nature. For example, the sound of harsh winds on the Mongolian steppe may have inspired early polyphonic singing traditions like Tuvaan throat singing. New technology developments now allow us to take this even further. Google’s Magenta team has been doing fantastic work for the last two years in the realm of art and AI. One of their coolest outputs to date is N-Synth, an AI-aided synthesizer that builds on traditions like spectral music in creating synthetic sounds humans have never heard before. We have just started playing with N-synth to see if we can generate sounds that precisely enhance the features correlated with frisson (acoustic roughness, harmonicity, tonal volume, etc.) and will share any breakthroughs we have.
But this is just a fun little side experiment; I think the best way to get breakthroughs from new tools like N-synth and Qbrio is to get them into the hands of up-and-coming music creators. Letting practitioners tinker with new tools almost always leads to unexpected but awesome applications.