¿ìè¶ÌÊÓÆµ

Songwriter AI emulates the Beatles with a little help from its friends

Many songwriters have honed their skills listening to the Beatles. Now AI has done the same by learning to write pop songs that mimic their style
The Beatles
On average people scored AI-generated Beatles songs 3 out of 5
dov makabaw sundry / Alamy

Thousands of songwriters have honed their skills listening to the Beatles – now there’s one more. An AI developed by a team at Snap, the company that owns the social network Snapchat, has learned to write pop songs in the style of the Fab Four.

Computers have been composing music for years. Many video games use computer generated soundtracks and AI-composed symphonies have been performed by live orchestras.

But classical or electronic music is easier for machines to produce because it can be more abstract. Pop songs like those written by the Beatles typically require a human ear for harmony, with catchy vocal melodies and chord progressions. Vocals, guitars and bass also tend to come together to play complex harmonies.

Yichao Zhou, at Snap and the University of California, Berkeley, and his colleagues took up the challenge with BandNet, a neural network that tries to mimic the genius of Lennon and McCartney. The results are more muzak than music but with a little tweaking, the software has come up with a handful of tunes that get your foot tapping.

The system was trained on 123 Beatles songs. Some of its early tunes contained bum notes because it had learned the Beatles’ trick of playing notes outside the key of the song. But unlike the Beatles BandNet could not tell when an odd noted sounded good and when it didn’t. To fix this the team restricted the software to only using notes within a given key.

BandNet does not generate its own song structures either. The type of AI algorithm used called a neural network has a short memory, which makes it hard for it to learn longer structures such as repeating verse-chorus-verse patterns.

Zhou and his colleagues asked a professional composer to score the content and structure of the system’s compositions out of 5. The composer gave the AI-generated tunes an overall average of 3.65 compared to 4.59 for the original Beatles songs.

The team also asked 17 volunteers to score BandNet’s tunes based on how similar they were to Beatles’ songs, how professional they sounded and how interesting they were. On average people gave BandNet just over 3 out of 5 across all categories.

BandNet won’t be topping the charts just yet but it shows that pop is a style of music that AIs can handle. The researchers think that software like BandNet could soon let anyone write their own Beatles-style song.

¸é±ð´Ú±ð°ù±ð²Ô³¦±ð:Ìý

Topics: Music