The guitar is in your hands. You want a slower beat, a cleaner sound, and harmony on the next chorus. Those are musical ideas. They ought to be things you can say.
That is the appeal of Muse, the musician’s assistant I have been developing with AI help. I want to be able to work on an arrangement in the language I use to think about it. Try a sound. Change the rhythm. Keep the combination that suits the song. The computer can deal with which control receives which instruction.
The project brings voice commands, chord charts, practice tools, and music controls together. It is still in development. The appealing possibility is already much more specific than “AI helps me practice”: talking an arrangement into shape while holding an instrument.
From “cleaner” to a sound
Muse’s documented voice flow starts with speech transcription. An AI classifier decides what kind of request it has received: change a chart, select a sound, adjust the tempo, control a loop, or answer a music question. The appropriate handler then performs the action. MIDI supplies the short messages that select sounds or operate musical controls.
The command vocabulary includes selecting a clean tone, changing tempo, turning harmony on, and starting or stopping a loop. Tone preferences can also be named and saved. Calling a sound “warm” is much closer to the way I want to remember it than remembering a preset number.
Imagine working out a stripped-back version of a song: choose a clean guitar sound, bring in a restrained rhythm, then try harmony. Each choice is a musical experiment. Does the arrangement need another layer, or did it sound better before the addition?
Save the arrangement with the song
Once a combination works, I want the song to remember it. Muse’s design includes saving rhythm selections into the OnSong chart and retaining song-specific tone preferences. The chart becomes more than the chords: it carries some of the preparation for my version of the song.
There is a useful division here. Muse is intended to help during preparation; OnSong carries the saved chart instructions into performance. I like that separation because experimenting with a sound and recalling a settled choice are different jobs. Onstage, the exciting variation should come from playing, not from the assistant reconsidering the arrangement.
Leave a space in the backing track
Another part of Muse works with stems: separate audio parts extracted from a recording. The documented workflow uses Demucs for separation, then offers backing-track mixing and export. Instead of only playing along with a complete mix, the idea is to leave room for the part I want to supply.
For practice, that changes the question from “Can I follow this?” to “Can I hold my part when it is my responsibility?” A backing mix with a part left out can create that space. The separation itself does not promise a pristine recording; I would still listen for what has been left behind or pulled into the wrong part.
I also like the opposite use: listening more closely to an individual part. An arrangement that sounds like one object becomes several decisions. What is the rhythm doing beneath the vocal? How much of the apparent fullness comes from the bass? Those are good questions to take back to a solo arrangement.
A tool for making my version
I do not want an assistant to decide my interpretation of a song. I want it to help me try one. A slower groove, a little harmony, a different guitar sound—and then the most useful musical instruction of all: keep that.