Stem Separation and What You Do With Each Half

"Split the stems" sounds like a single job. It's actually two different tools solving two different problems, and which one you need depends entirely on where the audio came from — something Versely generated, or something you found — and what you're planning to do once you have the pieces.

Two tools, not one, because the inputs are different

Stem separation — splitting a finished mix into its component parts as separate audio files — is the umbrella term, but Versely's chat surface actually exposes it as two distinct capabilities with two different input contracts.

Splitting a song into stems runs through separate_music_vocals, which operates on a track Versely already generated. It requires the source track's taskId and audioId from a prior generate_music or extend_music result — it's tied to a specific generation record, not an arbitrary audio file, and it splits that record into vocal and instrumental stems.

Isolating vocals from a track runs through isolate_audio, and it solves a related-sounding but genuinely different problem: extracting the vocal from any audio clip — a downloaded reference track, a voice memo, a video's existing soundtrack — by removing the background music and instrumentation around it. This is voice isolation in the strict sense — separating speech from everything else in a recording — and it takes an audio_url directly, with no generation record required, because there isn't necessarily a generation behind it at all.

The distinction isn't a technical footnote. "I made a song, now I want the pieces" and "I have some audio from somewhere, give me just the voice out of it" are different jobs with different starting conditions, and building one tool that tried to handle both would mean either forcing a taskId on audio that never had one, or losing the direct tie to a specific generation for audio that does. Two tools with two different input contracts is the correct shape for two different problems, not redundant coverage of the same one.

What the instrumental half is actually for

Once a generated track is split, the instrumental stem has a few concrete uses that aren't obvious from "here's the music minus the vocals":

• A clean bed under a voiceover-driven video. The instrumental carries the beat and energy of the original track without a competing vocal line fighting the spoken narration for the listener's attention — useful anywhere the generated song was the right vibe but the video needs a human voice doing the talking.
• A karaoke-style or lyric-caption format. Instrumental plus burned-in lyric captions is a distinct content shape from the original vocal track, and it only exists once the two halves are actually separated.
• Reworking just the instrumental for a different edit length. A video that needs a 15-second cut of a 3-minute generated song is often better served by trimming or extending the instrumental specifically, rather than re-cutting a mix where the vocal phrasing has to land in particular places.

What the vocal half is actually for

The isolated vocal stem is less obviously useful on first glance, but it earns its place in a few specific situations:

• Checking pronunciation and timing before committing to a full mix. Hearing the vocal alone makes phrasing and word-level timing issues audible in a way they aren't when buried in a full instrumental — useful as a QC step before a track goes into a final edit.
• Pulling an a cappella hook for a different edit. A striking vocal line from a generated track can be repurposed as its own short moment — an intro sting, a transition beat — independent of the instrumental it was generated alongside.
• Using the vocal stem's timing as a sync reference. When captions or on-screen text need to land precisely with sung or spoken phrasing, the isolated vocal — without instrumental noise interfering — is a cleaner reference track to sync against than the full mix.

Where isolate_audio does a genuinely different job

The separate case for isolate_audio shows up whenever the source audio didn't come from a Versely generation at all — a clip pulled in as a reference, an interview recording with music bleeding in from another room, a voice memo with background noise. This is a cleanup job on found audio, not a stem split on a generation: the goal is a clean voice track out of something messy, not two matched halves of something that was mixed cleanly to begin with.

A practical case where this matters: pulling a reference clip to study spoken delivery — pacing, emphasis, the shape of a sentence — before writing an original voiceover script in a similar register. Isolating the vocal from that reference makes the delivery itself easier to actually hear and study, separate from whatever music or noise happens to be in the source. The output is a study aid for your own original script, not a component meant to end up in a finished piece — a distinction worth keeping deliberately in mind, since the tool's job is cleanup, not licensing clearance for whatever the source clip actually was.

A real Versely walkthrough combining both

A realistic sequence that uses each tool for what it's actually for:

1. Generate a track with generate_music, and once it's rendered, split it with separate_music_vocals using the taskId and audioId from that generation — now there are two stems: vocal and instrumental.
2. Use the instrumental as the bed under a new video that needs narration rather than sung lyrics competing for attention, trimming or extending it to match the video's actual length.
3. Separately, and unrelated to the generated track, pull a reference clip with spoken delivery worth studying, and run it through isolate_audio using its audio_url — no taskId involved, because it isn't a Versely generation — to get a clean vocal track for studying pacing and phrasing before writing an original script in your own words.
4. Write and generate your own voiceover from that study, rather than routing the isolated reference vocal itself into a finished piece.

Two different tools, two different inputs, and — deliberately — two different outputs at the end: a reusable instrumental bed from your own generated track, and a clean study reference that informed an original script rather than became one.

The distinction that's easy to blur

The two tools sound similar enough — both involve pulling a vocal out of a mix — that it's tempting to treat them as interchangeable. They aren't. separate_music_vocals presumes you're the one who made the track and already have the generation record to prove it; isolate_audio presumes you don't, and just have a URL. Reaching for the wrong one either fails outright (no taskId exists for a clip you didn't generate) or gives you a stem split you didn't actually need (an instrumental half of a track you were only ever using as a reference, not as material). Knowing which situation you're in before picking the tool is most of the job.

The question worth asking before either tool: is this audio mine — a Versely generation with a record attached — or is it audio I found, with no generation history at all? That single question routes correctly almost every time, because it's really the same distinction the two tools were built around in the first place, just phrased as something you can ask yourself instead of a parameter you have to already know you need.