This paper reports a “musical Turing test”conducted at a live concert of algorithm-generated performances, where one group of participants were invited to rank the most human-like performance while knowing that one of the performances was by a human, and another group of participant were asked to do the same, but without knowing that there was a human performer on the program. The program consisted of five pieces from the classical/romantic period, played on a Disklavier. High quality music-expression algorithms were used to generate the algorithmic renditions. Regardless of the group, musical experience and a number of other factors, the subjects were unable to identify the human performer out of the five performances. The group that did not know there was a human performer had a wider range of votes compared to the group that did know. Furthermore, subjects were less confident of their answers when they knew that they were comparing human and computer-generated performances . On the contrary, if subjects believed they were only comparing computer-generated performances the task may have been less demanding. Findings suggest that computer algorithms are able to substitute for human performance, but the role of the physical presence of the performer (who was absent in this study) could be an area for further investigation.
Toward a musical Turing test for automatic music performance
RODA', ANTONIO;DE POLI, GIOVANNI;CANAZZA TARGON, SERGIO
2015
Abstract
This paper reports a “musical Turing test”conducted at a live concert of algorithm-generated performances, where one group of participants were invited to rank the most human-like performance while knowing that one of the performances was by a human, and another group of participant were asked to do the same, but without knowing that there was a human performer on the program. The program consisted of five pieces from the classical/romantic period, played on a Disklavier. High quality music-expression algorithms were used to generate the algorithmic renditions. Regardless of the group, musical experience and a number of other factors, the subjects were unable to identify the human performer out of the five performances. The group that did not know there was a human performer had a wider range of votes compared to the group that did know. Furthermore, subjects were less confident of their answers when they knew that they were comparing human and computer-generated performances . On the contrary, if subjects believed they were only comparing computer-generated performances the task may have been less demanding. Findings suggest that computer algorithms are able to substitute for human performance, but the role of the physical presence of the performer (who was absent in this study) could be an area for further investigation.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.