A machine that can detect a traffic accident by listening to the sound of screeching tyres, identify an industrial drill before it fails, or help musicians improve their compositions could change how artificial intelligence interacts with the world. At the Massachusetts Institute of Technology (MIT), Professor Paris Smaragdis is working towards that possibility by developing AI systems that can interpret sound, distinguish between overlapping audio sources and extract information that may escape human attention.In a profile published by MIT News, Smaragdis explains how his early interest in music and electronic instruments eventually led him to explore machine learning and computational audio. His research brings together music, engineering and artificial intelligence, with the broader aim of making machines better listeners rather than simply more capable processors of visual information.
From classical guitar to computer-generated sounds
Growing up in Greece, Smaragdis learnt classical guitar at an early age. His interest took a different direction when his father brought home a synthesiser, introducing him to the possibilities of creating entirely new sounds through technology.According to MIT News, this fascination eventually brought him to Berklee College of Music in Boston, where he explored computers and synthesised sound. He also developed an interest in jazz fusion and chose bass as his principal instrument.His work with electronic music tools led to a collaboration with Barry Vercoe at the MIT Media Lab. While pursuing graduate studies and later a postdoctoral position there, Smaragdis gradually shifted his focus from creating computer-generated music to understanding how computers could interpret sound.
Why teaching machines to listen matters
Smaragdis believes that machines need to learn how to recognise sounds rather than rely entirely on rules designed by humans. While conventional signal processing can handle specific audio problems, machine learning offers a way to train systems to identify patterns in complex sound environments.The applications extend well beyond music. Audio cues could help monitoring systems detect unusual events, identify mechanical faults or recognise moments of significance in a recording. For instance, a sudden rise in crowd noise during a football match could help identify an important moment without requiring a computer to analyse every frame of video.MIT News notes that Smaragdis explored such applications during his industry career, including work on systems that analysed audio accompanying video and helped users identify relevant content.
AI tools that support musicians
His research has also found practical applications in music production. At Adobe, Smaragdis established an audio research team that developed tools to help musicians and producers manipulate recordings without requiring extensive technical expertise.Later, at the University of Illinois Urbana-Champaign, his team developed neural network-based methods to separate individual sounds from mixed recordings. The technology proved useful beyond academic research, helping restore archival audio for the 2021 documentary series The Beatles: Get Back. The production’s sound-editing team subsequently won a 2022 Emmy Award.
Rethinking the future of audio AI
After 15 years at the University of Illinois, Smaragdis returned to MIT as a faculty member to explore emerging questions in audio research. His current work includes developing more flexible audio-processing systems, improving speech recognition and examining how large language models can contribute to sound analysis.As highlighted by MIT News, one of his priorities is to build tools that expand musicians’ capabilities rather than replace them. His group is also investigating how computers could provide useful feedback on music, helping artists understand and refine their work.For Smaragdis, the challenge is far from solved. Although AI has advanced rapidly, he believes researchers are still exploring how best to apply it to sound and music. The goal is not merely to make machines hear more accurately, but to help them understand what they hear and turn that understanding into something useful.
