It's a bit monotonous, but Irish folk music doesn't need complex melodies. Since it's well-played by human musicians, it sounds pretty okay, really. I wouldn't call it catchy.
*Maybe someone could get "Talk to Transformer" to write a music review of this GAN-generated music. If musicians can fool critics with AI, then critics can fool musicians, too.
TalkToTransformer:
How hard would it be to produce a plausible album of folk music from our system, folkrnn — a machine learning model trained on thousands of Irish folk music tunes? A model built to explore, interpret, and classify the way songwriters use words with meanings? A model built to predict the success of a recording (assuming there's one)? The answer is in a long, long way. Not only can you do this with existing algorithms, but you can learn how and why they do what they do — with amazing power.
But that doesn't mean we can get anywhere near this kind of machine learning without some technical innovation. That's why we've seen so many attempts at machine learning over the years, including many of the best and most elegant approaches to that task. And, as one of the authors of my first book suggests, there is a big difference between the approaches I've described in these terms, which are all fairly powerful, and approaches which are just as bad. I think it's time to look at how these approaches differ in two ways, and where they have in common.
A New Kind of Machine Learning
At some level, machine learning isn't very different from classical computer science.
The traditional approach was to build big models to describe the behavior of a system, based on data sets and mathematical formulas. To be able to do this a computer had to....
https://soundcloud.com/oconaillfamilyandfriends/02-the-drunken-landlady-gan
How hard would it be to produce a plausible album of folk music from our system, folkrnn — a machine learning model trained on thousands of Irish folk music tunes (https://github.com/IraKorshunova/folk-rnn)? We hired professional musician Daren Banarsë (http://www.darenbanarse.com) for this challenge, which resulted in the album, “Let’s Have Another Gan Ainm”. Of the 31 tunes on the album, 20 of them come from material curated by Banarsë from several volumes of tunes generated by our system (https://highnoongmt.wordpress.com/2018/01/05/volumes-1-20-of-folk-rnn-v1-transcriptions).
In Gaelic, “gan ainm” means “no name”, which is how each folkrnn tune on the album is designated. We let Banarsë be free in how he used the computer-generated material — in most cases he made minor alterations, but some are more substantial (for example combining material from two generated examples into one tune). We recorded the album in January 2018 in the Visconti Studio, Kingston University, with professional musicians trained in Irish traditional music. In March 2018 we sent the album out for review to a variety of places without revealing the true nature of the material.
The reviews the album received were very positive; it was not described as being unusual or uncharacteristic of this kind of folk music. One reviewer wrote, “[A] fine collection of beautifully-played tunes it is. While it includes some well-known titles such as ‘Lord Mayo’, ‘The Blackbird’ and ‘Toss The Feathers (II)’ … the unnamed jigs, reels and airs here thoroughly deserve their inclusion.” We publicly revealed the source of the album in August 2018. More information about this album can be found in our technical report: http://kth.diva-portal.org/smash/record.jsf?pid=diva2%3A1248565&dswid=7310.
“Let’s Have Another Gan Ainm” (2018) features the following musicians: Tad Sargent (bouzouki), Bryony Lemon (accordion), Grace Lemon (pipes), Daren Banarsë (melodica), Eimear McGeown (flute/whistle), Rob Webb (fiddle). This album is a deliverable of the project Sturm and Ben-Tal,“Engaging three user communities with applications and outcomes of computational music creativity” (funded by UK Arts and Humanities Research Council, grant no. AH/R004706/1), https://gtr.ukri.org/projects?ref=AH%2FR004706%2F1.
