Our world is changing so rapidly that technologies that until recently seemed impossible and amazing to mankind are now breaking into everyday life and are already considered the norm. At the recently held online technology exhibition CES 2021, a unique development was showcased that seemed interesting to Ojo.
Developers from MIT presented a creation to the public: a neural network capable of highlighting the sound of individual musical instruments. It got the name "PixelPlayer". The peculiarity of the development is that the neural network is capable of analyzing information using three algorithms simultaneously. PixelPlayer first examines the graphics of the image. Then it identifies the audio track, and then synchronizes the received data. That is, artificial intelligence compares the audio signal with each pixel of the video file, calculating which of the instruments is quieter than the others.
For an effective learning process, the program was provided with 714 videos (with a duration of 60 hours) with solo and orchestral concerts. As a result, the neural network identified individual audio tracks in the video recordings that it "heard" for the first time. According to scientists, the program is capable of recognizing up to 20 musical instruments: violin, tuba,
guitar, trumpet, accordion, cello, erhu, clarinet, flute, saxophone, xylophone and others. But this is not the limit if you provide additional information for training. It is worth noting that in compositions where there are instruments very similar in sound, the program sometimes makes mistakes.
The potential for practical use of the program by the broad masses is promising. Thanks to PixelPlayer, you can improve the sound in low-quality music recordings.
Earlier, OXO said:
German scientists printed the speakers as thick as a paper sheet