From Mozard to Botzard: When ma­chines write our mu­sic

Kuwait Times - - TECHNOLOGY -

PARIS: Ma­chines are al­ready tak­ing our jobs, will they soon be writ­ing our mu­sic too? Swiss re­searchers said Thurs­day they have de­vel­oped a com­puter al­go­rithm that can gen­er­ate brand-new tunes in dif­fer­ent mu­si­cal gen­res. The deep ar­ti­fi­cial com­poser, or DAC, “can pro­duce com­plete melodies, with a be­gin­ning and an end, that are com­pletely orig­i­nal,” said co-de­vel­oper Flo­rian Colombo of the EPFL re­search univer­sity in Lau­sanne, Switzer­land.

And the melodies are “quite agree­able to lis­ten to,” he said. The DAC pro­gram uses a form of ar­ti­fi­cial in­tel­li­gence known as “deep learn­ing” that works in a sim­i­lar way to the hu­man brain in mem­o­riz­ing ex­pe­ri­ences and learn­ing from them. It is a fast-grow­ing field, with more and more pos­si­bil­i­ties open­ing up as com­put­ers grow stronger and data­bases larger. The DAC sys­tem is trained to “lis­ten” to ex­ist­ing tunes to learn what works and what does not.

It teaches it­self to pre­dict the pitch and du­ra­tion of ev­ery note fol­low­ing an­other. Once it is ac­cu­rate at pre­dict­ing 50 per­cent of note pitches and 80 per­cent of note du­ra­tions in ex­ist­ing songs, the ma­chine’s train­ing is com­plete. Then starts the cre­ation. “The deep ar­ti­fi­cial com­poser builds a string of notes from be­gin­ning to end, in­clud­ing the very first note,” said an EPFL state­ment. It picks a fol­low-up note for each note played, based on the range of prob­a­bil­i­ties it mem­o­rized. But the pro­gram is taught not to pick the sin­gle-most prob­a­ble note-rather one of the many be­tween least and most prob­a­ble.

Can­not fool them

“An al­go­rithm that al­ways opts for most prob­a­ble note will just keep re­peat­ing the same thing,” said Colombo. In­stead, the pro­gram was de­signed to pro­duce “an in­fin­ity of dif­fer­ent tunes”. It may be ver­sa­tile, but the DAC is no Mozart. Ex­perts asked to lis­ten to DAC- and hu­man-com­posed melodies can still tell the dif­fer­ence. “It will be a while be­fore an al­go­rithm will fool con­nois­seurs of Ir­ish folk mu­sic,” said Colombo.

The re­searchers tested their pro­gram on Ir­ish and Ashke­nazi Jewish folk mu­sic. The DAC was able to iden­tify the genre, and avoid mix­ing the two. Any mu­si­cal style can be used, said the team. For the mo­ment, the DAC is lim­ited to sin­gle-in­stru­ment com­po­si­tions. Even­tu­ally, its de­vel­op­ers hope it will cre­ate a full or­ches­tra score. The goal, how­ever, is not to re­place fle­s­hand-blood hu­man com­posers. “It is more a tool that can be used to stim­u­late cre­ativ­ity, to aid the process of com­pos­ing” in pe­ri­ods of writer’s block, said Colombo.

He cited Mozart, who is ru­mored to have thrown dice to pick notes. DAC-com­posed mu­sic may even­tu­ally be used for jin­gles, but prob­a­bly never as “se­ri­ous mu­sic,” said Colombo. “A good com­poser, with in­no­va­tive ideas, will never be sup­planted by an al­go­rithm,” he said. “A com­poser puts some­thing of him­self in what he cre­ates, and that a ma­chine can­not do.” Sev­eral com­pa­nies, in­clud­ing Google, Sony and IBM are work­ing on sim­i­lar projects. —AFP

Newspapers in English

Newspapers from Kuwait

© PressReader. All rights reserved.