LALAL.AI Online Stem Separation

A web separation service that splits a track into vocals, backing and individual instruments after upload.

Model profile

Vendor: LALAL.AI
Form: 网页服务
Input: 音频 · 视频
Output: 分轨 · 人声
Licence: 上传内容的保留与使用、产出商用、各档位差异需按官方条款核对，训练数据未公开
Verified: 2026-10-30

Stems Sample Clearance Popular singing and the microphone

What it lets you finish

Split a finished song into vocals and backing, then take drums, bass, piano or guitar separately; make an instrumental for a cover; pull a clean drum or bass fragment for sampling; lift a vocal or a line of dialogue out of a live video; handle a client's rush job without installing anything.

Inputs and outputs

You upload audio or video, choose the parts to take, and the service returns stems you can preview and download, with format and bit depth chosen at download. Video is stripped to its audio first. What the free tier accepts and lets you download is as stated on the official pages, and paid use is metered by minute or by file. How long uploads are kept and whether they feed model improvement are stated in the terms and should be read before use.

Strengths

No install: a few steps in a browser, no command line.
Many part options beyond vocals and backing, including drums, bass and piano.
Works on a whole song, intros, breaks and endings included, without manual cutting.
Built-in voice cleanup helps make a stem usable; the exact options are listed on the official pages.
Results download straight into a DAW, which suits non-technical roles and one-off jobs.

Weaknesses and failure modes

Queues and limits: free use is restricted, so long programmes and batches mean repeated uploads with no control over timing.
Privacy: uploading sends material off the machine, which rules out unreleased work and confidential dialogue.
Residue: wet vocals keep their space, dense mixes lose low end, and drum transients round off.
Reproducibility: the server-side model changes, so the same material processed months apart may differ; keep the stems from the run you used.
Transcoding: uploading a lossy file means a second pass of loss, and sibilance and highs suffer first.
Metering: when billing is by the minute, repeated attempts cost money, so decide which parts you need and process once.

Terms, price and alternatives

Questions to check: do you still own what you upload? Is uploaded material used to train models? How long are files kept, and can you delete them? Does commercial release need extra permission? May you redistribute the stems to third parties?

Pricing runs on minute packages or plans, differing in length, number of parts and download formats; regional availability is as stated on the official pages.

Alternatives: Demucs when you need offline, batch and reproducible runs; iZotope RX when separation and repair belong in one project; Adobe Enhance when you only need dialogue cleaned and no music split.

Listen in the library

Onde Anda VocêC12:16A bossa arrangement with the voice close to guitar and bass, for checking residue in the vocal stem; a reference for this kind of output, not a generation by this model (C1)DetailsdindiC11:44A small-ensemble song, few instruments but plenty of reverb, where the vocal stem keeps its space; a reference for this kind of output, not a generation by this model (C1)Detailspor una cabezaC12:46A dance piece led by accordion and strings with a crowded midrange, for checking instrument separation; a reference for this kind of output, not a generation by this model (C1)Detailsbattle 03C11:12Dense game battle music with drums and brass, for checking whether transients survive; a reference for this kind of output, not a generation by this model (C1)Details

All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).

Sources

在线分轨的输入输出与档位，依服务官方页面核对，未抄原文
上传素材的隐私与版权处理，依服务条款与通识核对，未抄文本

Used for fact-checking only; all prose is original.