How Moises splits songs into different parts
If you’re a musician — or really, if you’ve ever listened to a song — Moises will feel a lot like magic.
This AI-powered marvel isolates the instruments in a recorded song, using machine learning models to separate out the vocals, guitar, bass, drums, and more. That means you can remove a guitar solo and supply your own. You can isolate the beat on a hip-hop track. You can strip out everything but the vocals to make any song a cappella. And you can change a song’s pitch, detect and display chords, and play along with a Smart Metronome.
“The new generation of machine learning models has enabled so much,” says Geraldo Ramos, Moises cofounder and CEO.
Moises
Available on: iPhone, iPad
Team size: 100
Based in: USA, Brazil
Awards: iPad App of the Year (2024), Apple Design Award finalist (2025)
Founded by a Brazil-based team with musical backgrounds — “We’re tech guys first, but we’ve always had a passion for music,” says Ramos — Moises was conceived and launched on the web over a single weekend in 2019. Less than a year later, Moises debuted on the App Store, instantly garnering tens of thousands of downloads. Today, Moises has more than 60 million users, everyone from seventh-graders practicing at home to music teachers to professional artists, producers, and vocal coaches. It's also localized into 33 languages.
We caught up with Ramos, cofounder and COO Eddie Hsu, and cofounder and Chief Design Officer Jardson Almeida to discuss machine learning models, unconventional time signatures, and recording at Abbey Road Studios.
Let’s start with the music: What’s everyone’s artistic background?
Ramos: I play drums, Jardson is a singer, and Eddie plays the violin. Eddie and I have known each other since kindergarten. He has a background in classical music and knows music theory, so he’s kind of the professional. And there are artists all throughout the company: One of our artist relations specialists graduated from the Berklee College of Music, and the ML team is almost entirely made up of artists. They love the math, but they also love the music.
How did this all begin?
Ramos: My journey started because I wanted to remove the drum tracks from recorded songs so I could play along. Before machine learning, that was basically impossible. You could equalize the audio or remove the bass frequencies, but it wasn’t that effective. It really only became a reality around 2019.
How did you turn that reality into an app?
Ramos: Moises was kind of a weekend hackathon project. We originally found an open-source model created by a research team in France who’d published simple code for song separation. There was no UI, no app, no nothing, but it worked better than anything else. I thought, “OK, I’ll create a UI over the weekend and see what happens.” I launched it on the web on Monday, and by the end of the week, 50,000 people had signed up. That’s when we decided to move to an actual business. That hackathon was in November 2019, and we launched our iOS app in late 2020. Things escalated quickly.
How did you respond to having that big of a hit?
Ramos: It was great, but it also made us realize that if we wanted to be a true AI company, we needed to create our own model. So before the app was even released, we’d developed our first proprietary model trained with our own data.
Where does all that data come from?
Ramos: It’s a mix of licensed music and tracks created by our own musicians and producers. On the licensing side, Eddie leads data acquisition, creation, and annotation — we actually have an internal iOS app just for annotating. That’s how we evaluate the tracks, see if one separation is better than another, and rank them.
Hsu: That’s one of our big differentiators — we do all the labeling for our models. So we’re stepping into the Apple ecosystem for that pretraining preparation.
And you record your own music too?
Hsu: We do! We have a whole catalog of music that the world has never heard. We once commissioned a session at Abbey Road to record some pieces for our data science team.
How do you commission recordings in a way that fills out your data gaps? What do you tell musicians to do?
Hsu: Here’s an example: We have a model that detects a song’s time signature. A lot of the data that we license – like pop songs — is in standard 4/4 time. But to improve our models, we need more diverse data, so we recently commissioned a number of recordings in 5/4 and 6/8 time. We’ve done this to improve chord detection too: Pop songs may have very simple chords, but if you’re playing jazz or bossa nova, things can be more complex, so we need more data to create a model that detects that complexity. And we’re committed to honoring artists and copyrights, so we have to be very diligent about how we commission these works.
How many instruments can be separated?
Ramos: There are two levels of separation. The first splits off basic tracks: guitars, drums, vocals, that sort of thing. The second goes deeper and can separate lead guitar from the rhythm guitar, or the snare from the hi-hat from the kick drum.
Hsu: When we started, we had presets for four tracks. Now there are more than 20.
Moises is great for students, but it’s also used by professional musicians and coaches. How do you ensure it‘s useful for different skill levels?
Hsu: I think it depends on the feature. Everyone is a learner, right? A student who’s just getting started with drums may need to slow a song down to learn it. But that goes for professionals too; we’ve heard from more than a few drummers who needed to learn something challenging and used Moises to slow tracks down until they got the parts just right. That’s where we realized, “OK, there’s a lot of overlap here.”
Originally published June 9, 2025
