Apps, Stems and AI for Songwriting at the Piano
What language models, stem splitters and harmony assistants do for songwriting at the piano, where the ear stays the referee, and what you keep doing yourself.
Ask a language model for a chorus line and you get what has to come back: heart, pain, a sky full of stars, grammatically clean and interchangeable. More useful is a request for twenty images around a theme instead of lines, with the clichés banned explicitly; the line itself then takes shape at the piano, because only there can you hear whether its stresses land on the beats. This pattern, tool for the material, human for the decision, runs through every category of apps and AI services now offered for songwriting, and the field changes month by month.
Tools sorted by task
What works without AI stays the foundation: a metronome app, the voice memo on the phone, MuseScore for notation and lead sheets, GarageBand or Reaper as a recorder for the demo. The danger is the perfect demo that keeps a song from ever being finished. Backing apps like iReal Pro or Band-in-a-Box, and Moises with its tempo and key controls, play any progression as a backing track in any key; for playing along that is fine, say to push right-hand voicings through all twelve keys, with the side effect that you end up having practiced chord changes and no song. Scaler, Captain Chords and Hooktheory with its two parts TheoryTab and Hookpad belong to the harmony assistants: they find progressions in real songs and tell you which chord follows most often in the statistics. You need to know that this chord is seldom the most interesting one before taking the suggestions seriously.
Claude, ChatGPT, Gemini and the other language models are useful for alternative rhymes, metaphor lists, translation, checking a theory question and analyzing a progression of your own; copyright keeps them from quoting anyone else's lyrics, and whatever they claim about the chords of a real song I verify at the piano before I believe it. Suno, Udio and similar generative audio services build a complete recording from a text. As a style sketch that is usable, as a release it is tricky: who owns the result is legally open, and the labels are already in court.
Stems: hearing the original's left hand
The category with the biggest promises is Moises, Demucs, Chord ai and Chordify, tools that transcribe and separate stems. They recognize chords and split a recording into vocals, piano, bass and drums. In isolation you hear what disappears in the mix: whether the bass holds the root while the piano plays an inversion, how thinly the verse is scored, where the right hand stays silent. I hardly use these tools all the same. Working out the bass by ear, or at least trying to, is harder and gives more than the finished track, because the way there trains the ear and the track only delivers the result. Chord recognition gets shaky exactly where piano accompaniment lives: as soon as inversions, seventh chords and a dense production enter the picture. What shows up as C may be a C/E with the third in the bass, or a Cmaj7 with no root in the right hand. The ear remains the referee, and your own degree analysis is the reference you hold the result against, not the other way round.
The order protects the craft
Hook and progression come first at the piano, with no screen. Everything that arrives from outside in this early phase ends up in the song rather than in the toolbox. Only then come the cross-checks: look the progression up in Hooktheory, see who has used it before and where it went from there; have the language model propose replacements for the third chord, then test every proposal with both hands; let a generative tool play back an arrangement concept, then build it yourself. Someone who understands scale degrees, modal interchange, melody and arrangement still writes a better song than any tool available today. The only new thing is the speed at which an idea becomes audible, gets compared and gets thrown out. Test more ideas that way and you write better; skip the idea that way and you get smoothly produced mediocrity. As theory I find all of this interesting; in practice I mostly want to play and make music myself, a hundred times rather with real people than with a model. That is the goal behind the practicing. The lesson contains a table with all six categories, examples, uses and cautions, plus two tasks: have your own chorus progression reharmonized three times and judge each version at the piano in the vocabulary of the course, and break a song into its stems, play along to the bass track and hold what you hear against your own degree analysis.
The course lets you hear it: sound examples, diagrams and trainers on this topic are in Lesson 29 · Tools: apps, stems, AI.
Open lesson 29 in the Piano courseSame topic on the guitar: AI Tools and Stems for Guitarists Who Write Their Own Songs