The UKRI CDT in Artificial Intelligence, Machine Learning and Advanced Computing (AIMLAC) aims at forming the next generation of AI innovators across a broad range of STEMM disciplines. The CDT provides advanced multi-disciplinary training in an inclusive, caring and open environment that nurture each individual student to achieve their full potential. Applications are encouraged from candidates from a diverse background that can positively contribute to the future of our society.
The UK Research and Innovation (UKRI) fully-funded scholarships cover the full cost of tuition fees, a UKRI standard stipend of £15,921 per annum and additional funding for training, research and conference expenses. The scholarships are open to UK and international candidates.
Closing date for applications is 12 February 2022. For further information on how to apply please click here and select the "UKRI CDT Scholarship in AIMLAC" tab.
The aim of this project is to train a neural net to complete Jigs or Reels based existing databases of Irish music starting with the well-known collection O’Neill’s 1001 Irish Gems. (This is a natural choice as there are pre-existing large databases of midi files.)
One example where this has been tried is in the AI Music Generation Challenge where the competition winning Irish jig was generated by neural net using “artificial critics” to select tunes according to on a metric and intervallic metric [B.L. Sturm, AIMC, 2021].
From an aesthetic point of view, the jig sounds unnatural and would never be played by traditional Irish musicians, and probably not even recognized as being Irish. The more natural approach, essentially the one followed by a human composer, would be to start off with a short motif (say 3 or 4 notes that sound good) then work out where this should go in a melody. This means that you have a subsequence and would want to work out where best to place it in a fixed structure, then fill in the remaining notes: in other words, to train the network to position a subsequence within a grid, then fill in the remaining elements.
The user input would be a short motif. This would repeat at various stages (mostly deterministic) and the real problem is to train the network to build the rest of the melody around it. From what I can see, most approaches involved getting the network to recognize patterns which are already well-known to musicians, rather than dealing with the real problem of extending a short motif to make a tune. RNN's and LSTM have been widely used, with emphasis on going both backwards and forwards over the sequences generated, however, we a central role would be given by the hierarchical structure [motif, 2-bars, 4-bar phrase (w/o cadences), 8-bar period (antecedent/consequent), sequence of periods]. This would be the novelty in the learning process.
Related work would include conversion of Welsh/Scottish traditional music to Irish form; learning from Scottish traditional music databases; categorization of tunes by region; comparison with original compositions.