AI music stems and audio inpainting: generating parts instead of finished songs
September 11, 2026
AI music stems are individual instrument tracks. Two methods produce them: separation, which pulls a finished mix apart, and generation, which creates each part directly from a prompt. Audio inpainting regenerates a marked region of a track while leaving everything around it intact. Stable Audio 3.0 supports both, generating stems directly and inpainting against numeric mask boundaries.
Key takeaways
Separation and generation both produce stems, but separation is limited to what a finished mix already contains while generation composes each part from a prompt.
Audio inpainting replaces a marked region of a track and preserves everything outside it.
Stable Audio 3.0 sets the edit region with two numbers, inpaint_mask_start_seconds and inpaint_mask_end_seconds, and accepts lists for both, so several regions can be regenerated in one pass.
Part-level output is what lets a generative model sit inside a DAW session rather than beside it.
What AI music stems are, and the two ways to get them
A stem is one instrument or vocal part exported as its own audio file. Producers work in stems because a mixed track is a decision frozen in place, and a set of stems is a set of decisions still open: the bass can be replaced, the drums re-compressed, the pad muted for eight bars. Two different operations put stems in a producer's hands, and the industry uses one word for both. Separation starts from a completed song and pulls it apart. Generation starts from a prompt and builds each part.
Stem separation: taking a finished mix apart
Separation is bounded by what is already in the recording. Suno's Advanced Stem Separation, as an example, offers three modes: Auto Split divides a track into up to 12 stems for 50 credits, Split from Mix extracts one chosen instrument plus a complement containing everything else at 10 credits per extraction, and Advanced Split picks from nearly 100 instruments at 10 credits per stem. Access is tiered, with Auto Split and Split from Mix on Pro and Advanced Split reserved for Premier, according to Suno's Advanced Stem Separation documentation.
The limits of the method show up in Suno's own documentation: requesting an instrument not actually present in the song may still consume credits without returning anything usable. ElevenLabs runs the same operation at coarser resolution, splitting a song into 2, 4, or 6 stems, priced at half a generation for two stems and a full generation for four, per ElevenLabs' own announcement.
Stem generation: prompting each part directly
Stable Audio 3.0 treats an isolated instrument as a generation target rather than a byproduct. Starting a prompt with the tag TrackType: Instrument raises the odds of getting a single clean part back, and Format: Duo asks for two instruments together. Genre and BPM tags narrow it further, so TrackType: Instrument, Format: Duo, a dynamic Latin drums and percussion track returns a percussion stem rather than a full arrangement with percussion in it, as set out in the Stable Audio 3.0 prompt guide.
Support varies by model. The medium and small-music checkpoints handle stems and solo instruments, and small-sfx handles sound effects only. Because the tag vocabulary comes from the metadata the models were trained on, drawn from Freesound and AudioSparx, descriptors a library would use tend to land: playing technique, mic perspective, room character, the sound of the gear. A generated stem carries no prior mix, so a producer can ask for a part that has never existed in any recording. What it gives up is the guarantee of fitting, covered further down.
Audio inpainting: regenerating a marked region without re-rendering the track
Audio inpainting in music production means replacing a section you have selected while the rest of the track stays byte-identical. The term arrived from signal processing, where it described reconstructing samples lost to scratched media or packet loss, usually across gaps measured in milliseconds, a problem still treated as signal reconstruction in the diffusion-based audio inpainting literature. That definition still dominates the search results and no longer describes what producers do with the tool.
The production use is deliberate rather than restorative. A chorus lands flat, a transition drags, a four-bar fill needs to be something else. Inpainting takes the marked region, generates new audio for it, and uses the surrounding material to keep the edges continuous. Stable Audio 3 runs this as one of three inference modes alongside text-to-audio and audio-to-audio editing, preserving the region outside the mask rather than re-rendering it, one of the three Stable Audio 3.0 inference modes. Audio-to-audio moves the whole file under init_noise_level control.
Mask boundaries, continuation, and multi-region edits
The edit region in Stable Audio 3.0 is two numbers. Setting inpaint_mask_start_seconds to 4.0 and inpaint_mask_end_seconds to 8.0 regenerates the fifth through eighth seconds and nothing else. The command line takes the same arguments as --inpaint-start and --inpaint-end, and the Gradio interface exposes them as mask start and mask end, documented in the Stable Audio 3.0 inpainting and continuation guide.
Two behaviors follow from that design. Continuation needs no separate feature: set the mask start to the current length of the file and the duration to something longer, and the model extends the clip. A 20-second bird call masked from 20s to 100s comes back as a 100-second recording. Multi-region editing needs no separate feature either, because both mask parameters accept lists. Passing [4.0, 16.0] and [8.0, 20.0] regenerates two disconnected windows in one pass. Multi-segment editing is named as a supported task in the Stable Audio 3.0 technical report, and it is closer to how anyone actually revises a track than fixing one spot and starting again.
Stability's own guidance on region size is worth following. Masking only a few seconds gives the model so much surrounding context that the output can return nearly identical to the original, so start with a large mask and tighten it on later passes. Prompts also need to be plausible against what surrounds them: a dubstep drop in a progressive house track will resolve, and a cat meow will not.
How to build a track from generated stems in a DAW
Structured composition is the workflow the two capabilities add up to. Fix the tempo and key first and state both in every prompt, because separately generated parts have no shared arrangement to lock them together. Generate the rhythm section, then the harmonic parts, then the lead, each as its own TrackType: Instrument call. Import, align, and mute what does not work. Where a part is right for 40 seconds and wrong for eight, mask the eight seconds and regenerate rather than rolling the whole stem again.
Stable Audio 3.0 reaches a session by several routes. The web app runs audio-to-audio and inpainting, exports stems as an experimental feature, and generates up to six minutes. The plugin runs the model inside Ableton, Logic, Pro Tools, and most other major DAWs, removing the export and reimport loop. Both surfaces are available through the Stable Audio web app and DAW plugin. For scripted work, the open weights run locally through the Python API, the CLI, or ComfyUI.
Stem and section-editing capabilities across generative audio tools
Several tools now produce stems and several now edit sections. The split runs between tools that perform those operations on a finished mix inside a hosted interface, and tools that generate the parts and expose the controls as parameters.
| Tool | How stems are produced | Section-level editing | Where it runs | Access to stems |
|---|---|---|---|---|
| Stable Audio 3.0 | Generated as isolated parts via prompt tags | Inpainting with numeric masks, multiple regions per call | Local, self-hosted, API, DAW plugin | Open weights on all four models |
| Suno | Separated from a finished song, up to 12 stems | Section regeneration in Studio | Hosted only | Paid plans, with Advanced Split at Premier tier |
| ElevenLabs Music | Separated from a finished song, 2, 4, or 6 stems | Inpainting API | Hosted only | Paid, priced against generation cost |
What generated stems cannot do
Stable Audio 3.0 does not produce intelligible vocals. The models sometimes generate vocal textures, useful as a pad or a hook and not lyrics. The Stable Audio 3.0 model overview states plainly that the models are not designed for speech or voice generation, so any project that needs a sung line needs a singer or a different tool.
Alignment is the other real constraint. Stems generated in separate calls are independent renders, so they will not sit together the way a separated multitrack does unless BPM and key are specified consistently and the results checked. Stem export in the web app carries an experimental label. Stable Audio 3.0 Large, the 2.7 billion parameter model, ships through the API and enterprise self-hosting rather than as open weights. The 1.4 billion parameter medium checkpoint needs a CUDA GPU, and the 433 million parameter small models run on CPU, per the Stable Audio 3.0 model table.
Frequently asked questions about AI stems and audio inpainting
What is the difference between AI stem separation and stem generation? Separation divides a finished song into its component parts, so the output is limited to what the mix already contains. Generation creates each part from a prompt, with no prior mix involved. Suno and ElevenLabs separate. Stable Audio 3.0 generates stems directly.
Will separately generated stems line up in my DAW? Not automatically. Each generation is an independent render with no shared arrangement, so specify the same BPM and key in every prompt and expect to nudge parts into place. Generating one longer master part first and prompting the rest against it works most reliably.
What is the difference between audio-to-audio and inpainting? Audio-to-audio transforms your whole file, with init_noise_level controlling how far the result travels from the source. Inpainting changes only the region between the mask boundaries you set.
Can I inpaint a track I recorded myself? Yes. Inpainting takes any seed audio, including your own recordings and earlier generations. Load the file, set the mask start and end in seconds, and add a prompt describing what should fill the region. Prompts that fit the surrounding material succeed more often.
How long a section can I replace with inpainting? No fixed limit applies below the model's maximum duration, 6 minutes 20 seconds on medium and 2 minutes on the small models. Short masks tend to return audio close to the original because the surrounding context dominates, so Stability recommends starting large and tightening on later passes.
Can I fix two different spots in one pass? Yes. Both mask parameters accept lists, so start values of 4.0 and 16.0 with end values of 8.0 and 20.0 regenerate two separate windows in a single generation. Non-contiguous editing in one call is documented in the Stable Audio 3 repository.
Do I need a GPU to generate stems locally? Not for the small models. Stable Audio 3 Small-Music runs on CPU at 433 million parameters and generates up to two minutes. The 1.4 billion parameter medium checkpoint needs a CUDA GPU and extends to 6 minutes 20 seconds with higher musicality.
Do I own stems and edits I generate with Stable Audio 3.0? Under the Stability AI Community License, you own your outputs and can distribute and commercialize them. Organizations above one million dollars in annual revenue need an Enterprise License, which adds legal indemnification. All models in the family were trained on fully licensed data. Full terms are in the Stability AI Community License agreement.
Last updated September 11, 2026
