Moises

For about thirty years, learning a song by ear meant fighting the recording. Drop the needle, rewind the tape, scrub the waveform, crank the EQ to dig a bassline out of a mix that was never designed to be taken apart. A generation of guitarists learned solos at half speed on a cassette deck with the pitch sliding a whole tone flat. Moises makes that entire ritual obsolete, and it does it from a phone, in about ninety seconds per song.

The app comes from Moises Systems, a team that built its business around music-focused machine learning rather than general-purpose AI wearing a music costume — a distinction that shows up in every corner of the product. The core is source separation: feed it a finished stereo track and it returns isolated stems for vocals, drums, bass, guitar, piano, strings and more. Everything else in the app is built on top of that primitive. Mute the guitar and there’s a band to play with. Solo the vocal and there’s an a cappella for study. Keep only the vocal-less mix and there’s a karaoke track that didn’t require a licensed backing version to exist.

Positionally, Moises sits in a spot that used to be occupied by four or five separate purchases: a slow-downer like Amazing Slow Downer or Anytune, a chord-detection service like Chordify, a stem tool like Lalal.ai or iZotope’s RX, and a metronome. It consolidates all of them into a single subscription — and consolidation is exactly what the market rewarded. A 4.7 store rating across tens of thousands of ratings isn’t the number of a novelty. That’s a utility working musicians have absorbed into their routine.

The competition has noticed. Apple shipped Stem Splitter inside Logic Pro. DJ software has neural stem separation baked in. Moises’ edge isn’t that it separates audio — it’s that it’s the only one built on the assumption of a musician holding an instrument, not a mouse.

• AI Source Separation: The engine. Uploaded or streamed tracks get decomposed into isolated stems — vocals, drums, bass, guitar, piano, strings, and other instruments depending on tier. Free covers vocals/drums/bass; the advanced multi-instrument models and Hi-Fi quality live behind Premium and Pro.
• The AI Mixer: Per-stem faders on a finished record. Drop the bass to zero and play the part. Pull the vocal down 6 dB instead of killing it outright when a guide is wanted but not a crutch. This is the feature that turns separation from a party trick into a rehearsal tool.
• Speed and Pitch, Decoupled: Slow a passage to 50% without the pitch sagging, or transpose the whole track to fit a vocal range or a capo position. Independent time-stretch and pitch-shift, done properly.
• Smart Section Looping: AI song-structure detection labels intro, verse, chorus and bridge, so looping the second chorus is two taps instead of a minute of dragging playhead handles.
• Automatic Key, BPM, Chords and Tabs: Chord charts scroll in time with the track. Generated guitar tabs give a starting point for the riff. BPM and key detection feed the metronome and the transposer.
• Count-In Click and Setlists: A metronome that counts in, and gig-day organization for the tracks actually needed on stage.
• DAW-Grade Export: Mixes and individual stems export in high quality, which is what pushes Moises from “practice app” into a legitimate first step in a production chain.

The Engine, and Where It Cracks

Honesty about what separation is: an inverse problem with no clean solution. A stereo mix is a sum. Reconstructing the addends from the total requires the model to hallucinate — intelligently, but hallucinate — what was there before the summing. That means artifacts are not a bug to be patched. They’re the physics of the task.

By design and reported results, Moises handles this better than almost anything at its price. On well-recorded material with clear separation in the arrangement — a rock trio, a singer-songwriter track, most modern pop — the stems should come out close to clean. Vocals retain body rather than the thin, phasey ghost that older masking-based tools produced. Drums keep transient snap. Bass, the easiest stem by a wide margin, is essentially a solved problem.

The cracks appear exactly where an audio engineer would predict. Dense mixes are where it struggles. Heavy walls of distorted guitar, cymbal wash, and long reverb tails are the enemy: reverb in particular is a spatial smear that belongs to no single instrument, so the model has to guess an owner, and it guesses wrong. A snare’s tail can migrate into the vocal stem. Mid-range instruments confuse it too — a clean electric guitar and an electric piano occupy overlapping territory, and the “guitar” stem can arrive with a piano stowaway. Solo a stem and push it hard and the watery smearing of the reconstruction gives itself away.

Here’s the thing: it doesn’t matter for the use case. When muting the bass to play along, minor bleed in the remaining stems is inaudible against the instrument in the room. Artifacts only become a problem when a stem is the star of the show — which is a production workflow, not a practice one. Judged on what it’s for, Moises is very good. Judged as a mastering-grade unmixer, it’s merely good.

Chords, Keys, and the Jazz Problem

The chord detection is the app’s most honest weak point. On diatonic pop, it’s reliable. Hand it modal jazz, extended harmony, or anything with secondary dominants and modal interchange and it flattens the world into major and minor triads. A ♭9 becomes a plain dominant. A modal vamp gets forced into a key center it doesn’t have. Key detection follows the same bias.

This is a known limitation of the whole field, not a Moises failure — but the app presents its analysis with more confidence than the analysis has earned. A visible confidence indicator would cost nothing and would stop an intermediate player from learning a chord chart that’s subtly wrong. Right now, the burden is on the user to know enough to catch the mistakes, which is precisely backwards from who this feature helps most.

The Loop That Actually Matters

The workflow is where Moises earns its reputation. Import a track. Wait for cloud processing. Land on the mixer. From there, everything needed is one gesture away: kill a stem, set a loop, slow it down, add the click, transpose it. That’s a full rehearsal setup, assembled in under two minutes, with no MIDI, no DAW, no rig.

Section detection is the underrated hero here. Precise looping has existed for decades, but it always required fiddly manual boundary-setting — the part of transcription practice that kills momentum. Being handed “verse 2” as a tappable object removes the friction between wanting to practice something and practicing it. That’s the difference between a tool admired and a tool opened every day.

The cloud dependency is the structural tax. Processing happens on servers, so there’s a wait, an upload cap on the free tier, and no offline separation. On a plane, Moises is a player for whatever was prepped in advance. Prep ahead of the gig or the flight — that’s the workaround, and it’s not a small caveat for touring musicians.

The design is restrained in a way that deserves credit. Music apps love skeuomorphism — fake knobs, fake wood, fake VU needles that convey nothing. Moises resists. The mixer reads as a mixer through layout and typography rather than pretending to be a physical console, which means it stays legible on a 6-inch screen where a photorealistic desk would collapse into mush.

Waveforms are clear, section markers are color-coded, and the touch targets on the faders are sized for actual thumbs. Scrolling chord charts stay locked to playback, which is a harder synchronization problem than it looks and one plenty of competitors fumble.

Performance should split along a predictable line. Playback is the light part: multi-stem real-time mixing with pitch and tempo manipulation is a nontrivial DSP load, but a job well within reach of modern phone hardware. Processing is the wait — and it’s server-side, so it depends on the connection and the queue. Free-tier processing sitting at the back of that queue reads as capacity management dressed as a tier difference.

iOS/iPadOS is where Moises is clearly happiest. The iPad build justifies its recognition: a larger canvas turns the mixer from a compromise into a proper control surface, with stems, chords and section markers visible simultaneously instead of fighting for the same vertical space. Propped on a music stand next to an amp, an iPad running Moises is functionally a rehearsal rig. Files integration makes getting audio in and stems out straightforward.

One warning that applies to both platforms and deserves more attention than the app gives it: do not practice through Bluetooth headphones. Wireless audio latency puts monitoring tens of milliseconds behind reality, which is fatal for timing work. Wired, or a low-latency setup, or the practice is training a player to play late.

Android carries the larger install base and a more fragmented hardware range than Apple’s. The app should behave well across that range because the heavy lifting happens in the cloud; a mid-range Android phone runs the mixer fine because it isn’t running the separation model. Background playback works. File handling is a little more variable across manufacturer skins, the usual Android tax.

The web app is the quiet third platform and the one to use for bulk work: batch imports and stem exports headed for a DAW are simply faster with a keyboard and a real file system.

Moises is that rare utility where the marquee technology and the actual daily value line up. Source separation is the headline, but the app’s real achievement is what it built around it — the loop, the click, the transposer, the setlist. Those are the unglamorous features that turn a demo into a habit, and the store rating suggests the habit stuck.

The frustration around the app isn’t about quality. It’s about a value proposition that keeps quietly moving. Capabilities that once sat in the free tier now sit behind Premium or Pro, and long-time users are right to be irritated when the app they adopted under one deal is now offering them a different one. That’s the pattern of a company that found its price ceiling by testing it against a loyal base — legitimate as business strategy, corrosive as goodwill.

Even so: the free tier alone still does something that was expensive and slow five years ago and impossible fifteen years ago, and it does it in a pocket. For a practicing musician, this earns a spot on a phone ahead of most apps — with eyes open. Budget for the subscription, wire the headphones, and don’t trust the chord chart on a Coltrane record.