Stems for a wide range of guitar sounds

AudioShake Research
April 2, 2024

We’re a team full of guitar players, making it an instrument and stem we care deeply about! 

AudioShake’s guitar stems have fueled a wide range of creative uses since their original launch, including “guitar karaoke” with Green Day–in which we worked with the band to separate out Billie Joel’s performance from “2000 Light Years Away” and let users become the band’s guitarist. They’ve also powered many projects across major and independent labels, including immersive, AR/VR, sync, and more. 

We recently released a new guitar separation model, and as with our new piano stems, these improvements were so great that we wanted to capture what challenges we set out to tackle, as well as the perceptual and quantitative improvements we made. For our latest models, our team built entirely new models and data pipelines to create even higher fidelity isolations and improve sustain for guitars across styles, genres, and pedal effects. 

“Our new guitar model harnesses the most advanced data pipelines we've ever constructed, engineered to simulate an extensive array of guitars, amplifiers, and, most notably, effect pedals. These machine learning models are intricately designed to detect the broad amplitude range of guitar sounds, encompassing everything from the crisp, clear tones of western guitars to the heavy, saturated distortion of a “Big Muff” pedal.” – Fabian-Robert Stotter, Head of Research at AudioShake

The challenge with isolating guitar stems is that guitars can almost sound like anything these days thanks to pedals and effects, creating an infinite range of possible sounds. 

“This advancement signifies a monumental leap in our ability to capture and reproduce the dynamic essence of guitar music, ensuring that every nuance and detail is preserved and accurately represented.” – Fabian-Robert Stotter, Head of Research at AudioShake

Perhaps the best way to understand the differences between our old and new guitar stems is to hear the difference: