The problem
StemDeck can tell you where every beat is, but nothing about where the song
goes. To loop the second chorus, or drop into the bridge, or hand a bandmate a
timestamp, you have to find the boundary by ear and drag a marker there
yourself, for every song, every time.
The sections bar already exists and is already editable. What is missing is
anything to put in it.
What makes this harder than it sounds
- Model output is untrusted input. Whatever produces the suggestion is a
third-party model. Its timestamps, its labels, and its structure all have to
be validated before they touch a job, and a malformed result has to leave the
separation exactly as it was.
- It costs real time. Inference is a CPU pass measured in minutes. A user
who only wanted stems should not silently pay for it, which means it needs to
be a visible choice rather than an always-on cost.
- The stems are the wrong shape. StemDeck separates six stems; the models
in this space expect four. Mapping between them without running a second
separation is a constraint on any solution.
- Labels will be wrong sometimes. Functional names are not ground truth
even when the boundaries are right, so whatever ships has to present them
honestly and stay editable rather than claiming more confidence than it has.
- Failure must be non-fatal. This is the last stage of a pipeline that has
already done the expensive work. Nothing here may cost a user their
separation, and cancellation has to stay responsive throughout.
The problem
StemDeck can tell you where every beat is, but nothing about where the song
goes. To loop the second chorus, or drop into the bridge, or hand a bandmate a
timestamp, you have to find the boundary by ear and drag a marker there
yourself, for every song, every time.
The sections bar already exists and is already editable. What is missing is
anything to put in it.
What makes this harder than it sounds
third-party model. Its timestamps, its labels, and its structure all have to
be validated before they touch a job, and a malformed result has to leave the
separation exactly as it was.
who only wanted stems should not silently pay for it, which means it needs to
be a visible choice rather than an always-on cost.
in this space expect four. Mapping between them without running a second
separation is a constraint on any solution.
even when the boundaries are right, so whatever ships has to present them
honestly and stay editable rather than claiming more confidence than it has.
already done the expensive work. Nothing here may cost a user their
separation, and cancellation has to stay responsive throughout.