Rise of the Music Machines

May 16, 2023·50 min·4 clips

Ten musicians spent 38 days describing 5,500 music clips so Google could build an AI that creates music from words.

Google began with drudgery. Ten musicians worked through 5,500 YouTube clips and turned each musical fragment into prose a model could match against sound. The numbers are ridiculous: 38 straight days of listening, 92 hours per worker, and 370,000 words of careful description. That labor is the premise hiding in plain sight. Before the episode judges the outputs, it shows how human musical perception gets flattened into training data for a machine. Then comes the listening test. The host and another speaker try period styles and genres, and the model gives back sounds that resemble public memories more than full histories. The 2000s example lands with a thud. It sounds like indie-sleaze, while the other speaker notes that bling rap and hip hop clubs would tell a different story. Grunge does a little better. Even there, the judgment stays careful, placing the track near alt-rock radio instead of treating it like a definitive reconstruction. Mood music changes the stakes. Chill out and down tempo feel less like tests of genius and more like clues to where automated music can make money. Nobody needs a hit there. The likely market is ambient background audio for YouTube channels, exercise videos, and steady lo-fi streams where attention is meant to stay loose. That is where the episode gets sharp. The machine does not have to beat musicians at expression if it can fill every low-stakes slot cheaply and endlessly. The ending looks beyond the test, using links to Google Music LM, OpenAI Jukebox, and related tools as a map of systems already arriving.

As heard by us

The 38 days of human listening behind Google's AI music machine.

Ten musicians spent 38 days annotating 5,500 YouTube clips -- 370,000 words of description -- before MusicLM could generate a single bar. That origin story is what gives the episode its grip, and Hi-Phi Nation earns it, using the bureaucratic setup to push toward a harder…

Read the full review in PlayNext →

Why you'd press play

Before MusicLM could generate a note, ten humans wrote 370,000 words describing music for Google.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Hi-Phi Nation