Rise of the Music Machines

May 16, 2023·53 min·4 clips

Ten musicians spent 38 days describing 5,500 music clips for Google's AI to learn how to generate new music from words.

A tedious listening job sets up the question. Google gives ten musicians 5,500 clips from YouTube and asks them to describe, in ordinary words, what each fragment sounds like. They do it. Together, they listen to 38 straight days of music and write 370,000 words linking genres, vocals, atmospheres, and possible uses. That becomes the machine's lesson: training material for a deep learning model that has to learn which words match which musical sounds. The episode keeps things concrete. One R&B soul remix turns into a cluster of tags: laid-back vocal, auto-tuned voice, groovy atmosphere, sitcom soundtrack. Then the messy part arrives. The conversation moves through indie sleaze, bling rap, grunge, alt rock radio, and the problem of asking a model to land on a style people still argue about. Genre is never just filing. It carries audience, race, chart systems, nostalgia, and fights over who gets included under a musical name. The bigger market may be quieter. The guest argues that fully generated music may matter less as pop competition than as an endless supply of mood and background sound. That feels plausible. Chill out, down tempo, lo-fi study beats, exercise videos, podcasts, and YouTube channels all need music at a scale ordinary licensing has trouble serving. The risk is plain too. Once music becomes something you can prompt on demand, pressure falls on the less visible workers who already make functional sound. The episode stays clear-eyed about the novelty. Music machines may not replace the hitmaker first. They may simply fill every unattended corner where music is needed but barely heard.

As heard by us

A podcast traces how ten musicians' labor gave Google the ability to make music without them.

Before Google's MusicLM could turn a text prompt into a track, ten musicians spent 38 straight days listening to 5,500 YouTube clips and writing plain-English descriptions -- 370,000 words of human attention that the model then learned from.

Read the full review in PlayNext →

Why you'd press play

The receipt for AI music is longer than you think: 10 humans, 38 days, 370,000 words.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Hi-Phi Nation