Month by month, we continue to witness the unfolding of new possibilities within the realm of generative AI. Demonstrating this potential once again, Google has collaborated with researchers from Japan to pioneer a remarkable achievement: the generation of music derived from human brain activity. This groundbreaking endeavor involves capturing brain signals through functional magnetic resonance imaging (fMRI) and subsequently reconstructing them using Google’s innovative MusicLM music generation model.
Outlined in a research paper titled “Brain2Music: Reconstructing Music from Human Brain Activity” (via ZDNet), Google elaborates on their latest breakthrough. They teamed up with Japanese researchers to embark on an extraordinary journey of creativity. The venture involved selecting random 15-second segments from a pool of 540 music pieces spanning ten distinct genres. These snippets were then played for five participants through MRI-compatible insert earphones, all the while their brain activity was meticulously scanned.
The wealth of data collected was channeled into Google’s MusicLM, which skillfully utilized the information to foresee and rebuild the very essence of the music to which the human subjects had been exposed. In the culmination of this intricate process, the generated music exhibited intriguing parallels to the original compositions on a semantic level.
For those intrigued to experience the fruits of this innovative endeavor firsthand, Google has thoughtfully assembled a webpage dedicated to the study. On this platform, visitors can explore not only the initial musical stimuli but also the remarkable reconstructions that emerged from decoding human brain activity.
The research team also identified three key factors that contribute to the limitations in the quality of AI-generated music:
- The fMRI data contains temporal and spatial gaps: The regions observed in the fMRI data are relatively large (2×2×2mm3), significantly exceeding the size of human neurons by many orders of magnitude.
2. Music embeddings lack comprehensive information: The music embeddings used for music reconstruction (utilizing MuLan) condense ten seconds of music into just 128 numerical values, which results in a loss of detailed musical information.
3. Constraints of the music generation system: An assessment of MusicLM revealed areas for enhancement, both in terms of how it interprets text prompts and in the fidelity of the audio it produces.
As is the case with most advancements, there is ongoing work to be done in this field. However, AI has demonstrated its ability to skillfully leverage brain activity to recreate auditory experiences. Nevertheless, if privacy concerns arise—worries about someone scanning your brainwaves and accessing your thoughts—rest assured that current safeguards are in place. Volunteers involved in the study spent extended periods within large fMRI scanners, making such scenarios unlikely. So, for now, you can steer clear of large fMRI scanners when planning your first dates.