PixelOkay
Tools

AUDIO STUDIO / NO AI REQUIRED

Audio Stem Separator

Isolate center/side audio, frequency bands or harmonic/transient layers without AI. Optionally try genuine neural instrument separation. Preview, remix and download WAV/ZIP locally.

◎ Private processing◌ No-AI DSP by default↓ WAV & ZIP exports
01

Choose your audio

Drop an audio file hereor click to browse your deviceWAV, MP3, M4A, FLAC, OGG and other browser-decodable formats · Up to 80 MB

Project JSON stores settings only, not audio. Reattach your original file after reopening.

02

Choose processing

Processing mode

DSP uses acoustic properties; center/side output is NOT clean vocals, and frequency bands are NOT separated instruments.

Memory use can be multiple gigabytes during AI inference, especially in CPU mode.

Add an audio file to begin.

How it works

1Import a track

Choose a short segment of audio your browser can decode.

2Choose DSP or AI

No-AI DSP isolates center/sides, bands or harmonic/transient layers. AI is optional.

3Listen and export

Solo or mute stems, create a new mix, and save WAV files or a ZIP.

Frequently asked questions

Does PixelOkay upload my audio?

No. Decoding, AI inference and export happen in your browser. Downloading the AI runtime and model weights requires internet access on first use. The hosting site may receive ordinary page traffic.

Why is there a model download?

The model contains trained weights that identify individual instruments. SCNet is approximately 13 MB; SCNet Large is approximately 45 MB. Additional runtime files also download, and inference can require several gigabytes of memory. Browser caching depends on available storage.

Which instruments are included?

No-AI tracks represent center/sides, bands, or harmonic/transient components. Optional AI requests vocals, drums, bass, and other; model accuracy is not certified.

Why does separation take so long?

Neural separation uses significant computing power and memory. A GPU can help if the browser supports WebGPU. CPU mode may take several minutes, especially for longer clips.

Using Audio Stem Separator

Isolate real acoustic components without AI or optionally request neural stems. Audition and export real reviewed audio tracks.

Workflow

  1. Open supported audio within the source and duration limits
  2. choose a range and process locally using DSP, or explicitly consent to optional model assets
  3. audition the four stems and rebalance or export reviewed stems or mixes.

Example

Select a short music excerpt, isolate center/sides or four frequency bands, audition each processed layer and the reconstructed mix, then export only outputs that play correctly for your use.

Limits and data handling

Local DSP is available without a model. Real optional SCNet inference and model distribution are not verified. AI mode can download an ESM runtime and model weights from external providers after explicit consent. Model files may require several gigabytes of memory despite their smaller download size. Source is limited to 80 MiB and seven minutes; selected ranges are capped at 180 seconds on desktop and 90 seconds on small screens. No synthetic fallback is presented as real separation. Review provider rights, CSP/CORS, privacy and actual outputs before relying on results.

FAQs

Does separation work offline on first use?

No. No-AI DSP runs offline after the static site files are installed. Optional AI downloads a runtime and model assets after consent; actual inference remains unverified.

Are the four stems guaranteed to be clean or independent?

No. Listen to each result and the mix; separation quality depends on the model, source and device.

Related tools