Researchers in the UK have created an artificial intelligence capable of writing Irish traditional tunes which can be indistinguishable from the real thing.
The AI has already written over 100,000 folks songs, most of which are yet to the played by human musicians. The program was written by researchers from Queen Mary University and Kingston University in London.
The creators are now staging real life concerts to evaluate the quality of the music which is generated by a program called “folk-rnn”.
“We must remember that notes on paper are not yet music - it needs the performers who bring so much into it. So, we are working with professional musicians to see if some of the tunes produced by the program are good enough to make music from. Sometimes, it requires only a little tweaking; other times, it requires a lot,” said Lecturer of Digital Media at Queen Mary, Bob L Sturm.
The AI works by using machine learning to pick up patterns from human generated songs which are fed into it. To date over 23,000 music transcriptions, many of which are of traditional dance music from Ireland and the UK, have been fed into the program.
“We can then ask the program to generate new tunes, or we can ‘seed’ the system with a a few notes and ask it to autocomplete the rest,” Mr Sturm said. “The result is a computer that can produce music the world has probably never heard before, but which often sounds good.”
Like all songwriters the AI has been known to produce a more than a few duds. Sturm said the feedback from musicians at the moment is that only one in five of the songs are decent.
The creators focused on Irish traditional music because a large number of existing compositions are available online that can be fed into the program.
The presence of many Irish trad musicians in London was also a factor.
“In the UK we are surrounded by many practitioners of ‘session’ music. Their expertise in this music allows us to understand better what our system gets right but also, more importantly, what it doesn’t get about this music.”
Even when the program gets it wrong it can still help musicians, Mr Sturm said.
“Interesting examples arise when the program does something really weird and uncharacteristic of the style it has learned. These are failures in terms of learning the style but can be an interesting starting point for creation,” Kingston College senior lecturer and composer Oded Ben-Tal added.
The researchers are eager to point out they are not trying to replace human composers. They say folk-rnn can be used as a composition assistant to help compose new melodies or to teach students.
A beginner or a student might use the system to generate ideas.
“So far we trained the system in one musical style but we can also adjust our program to fit other styles as well, such as melodies from the renaissance, or atonal music.”
We are looking at how these technologies can augment one’s creativity.”
