Analytics Stories

CHAPTER 57Can Analytics Recognize, Predict, or Write a Hit Song?

I love trivia. Here's a great music trivia question. Of all albums recorded between 2000 and 2009, which album by a solo artist sold the most copies in the United States? The answer is Norah Jones's debut album Come Away With Me, which sold 11.02 million copies (rateyourmusic.com/list/abyss89/the_100_biggest_selling_albums_of_the_2000s_ _usa_/). When Jones recorded this album, she was unknown in the music world, but the computer program Hit Song Science used artificial intelligence (AI) to accurately predict that the album would be a huge hit (www.nytimes.com/2003/12/14/magazine/2003-the-3rd-annual-year-in-ideas-hit-song-science.html).

The key to using computers to recognize, predict, or write a hit song is to create an acoustic fingerprint of a song. Stripped to its essence, a song is a sound wave. The acoustic fingerprint of a song (or any sound) identifies at some (or all) points in time many attributes of a song, including

• Beats per minute (tempo).
• Frequency. For example, the top note on a guitar is 1,000 cycles per second, whereas a typical adult female singer sings a frequency between 165 and 255 cycles per second.
• During any part of a song, different frequencies (a guitar and a singer, for example) are combined, so you need to know at different points in time the amplitude of the song in different frequency ranges. Spectral flatness is used to measure at a point in time whether a song's sound is more like ...