It’s the moment everyone’s been waiting for all year: Spotify Wrapped time. Though it launched later than initially anticipated, the annual personalized wrap-up is available for subscribers to see where their music tastes landed for the year. What’s also available is a customized podcast courtesy of Google’s NotebookLM, featuring two AI hosts who struggle to pronounce any song or artist that is not of the Anglo-Saxon variety.
Artificial intelligence isn’t perfect. Even with the most convincing demonstrations at developer conferences, AI text-to-speech always sounds slightly forced compared to how humans talk. The best example is Spotify’s Wrapped AI podcast, which resembles Google’s demonstration earlier this year at Google I/O. The hosts are certainly convincing, but the longer you listen to the synthetically concocted podcast, the more obvious it becomes that you’re listening to a pair of robots recapping your listening stats.
The AI-produced podcast starts standard. It sounds like me and my podcasting partner on an episode of our show. First, the hosts highlight the longest day you listened to Spotify during the year. It picked up on a particularly emotional day when I was listening to hours of sad music. The AI hosts safely wondered aloud if it was because I was on a road trip that day. (I wasn’t. I was crying!) This is an opportunity where the AI could have scaled back on making assumptions. Like a friend who isn’t thinking about the impact of their words, the AI triggered me to harken back to a day of significant discomfort. Rude behavior!
Then, the AI hosts read off your most-played artists and songs of the year. I would still prefer to hear an actual, live human attempt to pronounce the name of my favorite Dutch artist, Joost Klein, than Google’s NotebookLM. His name is pronounced “Yost,” but the podcast instead refers to him as “juiced.” It’s even worse when they try to pronounce the Dutch songs. My most played song this year is called “Wachtmuziek,” and Google’s NotebookLM hosts pronounce it exactly as you might attempt to read it as a non-Dutch speaker, without the hawk-tuah that’s enunciated with the “cht”.
Spanish speakers will also find that Google’s NotebookLM struggles with the Latin language. My editor shared his personalized podcast with me, and I cringe whenever they attempt to identify a song by Bad Bunny. Don’t get me started on the butchered pronunciation of “ray-gay-ton,” which is not how reggaeton is said out loud. I would argue that uttering the word with the appropriate intonation is the genre. The cringe factor is high with this one.
This leads me to ask: Isn’t this something AI should pick up on? The music is categorized under the language in which it is presented. I imagine a bot could be programmed to pull in from that data and adjust accordingly. To the credit of its developers, Google covered its bases by mentioning that its AI hosts occasionally mispronounce words. But it takes the magic out of the AI when you realize the bot is just as flawed at failing to practice pronunciation as the rest of us.
Spotify’s Wrapped AI podcast with Google prowess is available for Free and Premium users in the U.S., U.K., Australia, New Zealand, Canada, Ireland, and Sweden for a limited time.
Read the full article here
