If you’ve watched the Nvidia GTC keynote speech, you are well aware that artificial intelligence can be used to create fantastic music. The soundtrack in the video was composed by AIVA , which is a company that uses artificial intelligence to create music.
In this article, I will examine AIVA’s music composition tool in detail. I am a machine learning engineer in the field of AI music composition, and I will offer my insight and experience with AIVA's tools.
Summary of the Tool
AIVA stands for Artificial Intelligence Virtual Artist. From the company website, AIVA is “artificial intelligence composing emotional soundtrack music.” It functions as a “creative assistant,” to help songwriters with their music production.
Since this tool is designed to create soundtrack music, this can be a great tool for composers who work in the film, TV, or video game industry. The tool can also be used by songwriters who are interested in writing pop songs, since it can also generate songs in the pop, rock, or electronic genre.
There is a 3 tier pricing structure with AIVA. The first tier is free, but all copyrights for the generated music belong to AIVA. The remaining tiers are monthly subscriptions, with different licensing and copyright options. The free option is a great way to try out the tool.
To create a track, the user is presented with a list of options of a preset style. Some of the styles include cinematic film scores, pop, and electronic music.
Each track takes around 1 minute to generate. The generated music has repeated melodic motifs and chord progressions, which indicates that the AI produces music that is similar to human composed music. As the melodic motifs change throughout the song, the structure of the motifs remains consistent. Overall, the generated music sounds very convincing and musical, as if the music were composed by a human.
Most of the generated music also contains full instrumentation, such as drums, guitars, and orchestral instruments. This is great for musicians who want to get a head start on their project. It also serves as a realistic representation of the final product.
The generated music can be downloaded as audio files (WAV, MP3) or MIDI files.
I’ve been provided with the following resources from AIVA’s customer development team :
AIVA’s initial training dataset consists of “30,000 of history's greatest scores.” These scores are mostly from classical music composers such as Beethoven and Mozart. More recently, they’ve increased their dataset to include rock and pop songs, with modern verse and chorus song structures.
Machine Learning Architecture
AIVA’s machine learning architecture appears to be some kind of autoregressive model, such as a RNN. They also provide checkers to check for plagiarism and the quality of the generated music.
If you’re curious about AI generated music, sign up for a free account. The tool is fun to play with, and generates convincingly human-like music.
Although there is not a lot of public information about the technical details of AIVA, the quality of AIVA’s music demonstrates the capabilities of the technology. I believe that we will be hearing a lot more of AIVA's music in the future, especially as soundtracks for movies, TV shows, and video games.
Special thanks to AIVA’s customer development team for providing the resources for AIVA's technical details.
About the Author
Wayne Cheng is the founder and machine learning engineer at Audoir. His focus is on the use of generative deep learning to build songwriting tools. Prior to starting Audoir, he worked as a hardware design engineer for Silicon Valley startups, and an audio engineer for creative organizations. He has a MSEE from UC Davis, and a Music Technology degree from Foothill College.