music, language & machine learning: The Text Score Dataset 1.0
![FOR A DAY: For a day, be a situationist provocateur. Try to convince as many people as possible that the moon is actually yoghurt. Return every question with the phrase "I'm trying to have a praxis." [Visualisation of performance by VQGAN+CLIP]](https://images.squarespace-cdn.com/content/v1/5300fd50e4b008058a48170a/1627638558097-5RIO40J5SAJVZI38NR7V/IV+For+a+day%2C+be+a+situationist+provocateur.+Try+to+convince+as+many+as+possible+that+the+moon+is+actually+yoghurt.+Return+every+question+with+the+phrase+%2522I%27m+trying+to+have+a+praxis.%2522.png)
FOR A DAY: For a day, be a situationist provocateur. Try to convince as many people as possible that the moon is actually yoghurt. Return every question with the phrase "I'm trying to have a praxis." [Visualisation of performance by VQGAN+CLIP]
Since 2017 Jennifer Walshe has been engaged in a quixotic enterprise, the creation of a massive – one might say definitive – dataset of text scores. Walshe’s Text Score Dataset 1.0 now comprises over 3,000 text scores, running to almost half a million words, ranging from Fluxus event scores to compositions written in the last year. Many of these scores were painstakingly transcribed by Ragnar Árni Ólafsson, who over the last four years has come to know this territory in a way very few other people on the planet do.
Jennifer Walshe created the Text Score Dataset 1.0 in order to be able to use it as training material for Machine Learning algorithms, so that new generations of text scores could be created. Commissioned by PRiSM Centre at the Royal Northern College of Music, over the last months she and Ólafsson have been working with David DeRoure of the University of Oxford’s Department of Engineering Science on the first generation of outputs.