service worker
The only context that can mint a tabCapture stream.
One tab. Six stems. Nothing leaves.
A free, open-source Chrome extension. Open a YouTube video, press the shortcut, and a six-channel mixing deck appears inside the page. A neural network runs on your own GPU and pulls the audio apart in real time — vocals, drums, bass, other, guitar, piano. Mute the singer. Solo the drums. Slow it to 0.75× without changing the key.
CtrlShift9
⌃⇧9 on macOS
Install it
- Model
- htdemucs_6s
- Runtime
- ONNX Runtime Web · WebGPU
- Latency
- ~3.4 s
- Uploads
- 0
- Network requests, lifetime
- 1
02Separation
One thing,
taken apart into six.
The tab plays a single stereo pair. The deck gives you six, each one on its own fader, mute, and solo — and a seventh passthrough lane so you can hear what the model was given.
In — one stereo pair, 44100 Hz
Out — six stems
- 01vocals1 · ⇧1 · ⌥1
- 02drums2 · ⇧2 · ⌥2
- 03bass3 · ⇧3 · ⌥3
- 04other4 · ⇧4 · ⌥4
- 05guitar5 · ⇧5 · ⌥5
- 06piano6 · ⇧6 · ⌥6
The six colours come from the deck's own tokens.css. They were chosen to
stay separable under deuteranopia and protanopia in combination with a fixed
position, a number, and a text label. Colour is never the only thing carrying the
state — here or in the deck.
03The surface
Six faders,
and no audio at all.
This is the extension's control surface, rebuilt here so you can put your hands on it before you install anything. The faders, the mutes, the solos and the keys are the real ones. The meters are generated. Nothing on this page makes a sound.
Six-channel deck
Demonstration · synthetic meters · no audio
VOCALS
LIVE
DRUMS
LIVE
BASS
LIVE
OTHER
LIVE
GUITAR
LIVE
PIANO
LIVE
Master
Speed · key-locked
Transpose
This is a demonstration, not a player. It is a working model of the extension's control surface, driven by generated meter data. It plays no audio, and it never will on this page — there is no audio on this site at all. Drag a fader, or press 1–6 to mute and Shift+1–6 to solo: the keys are bound right here.
Each channel spells its own state out — LIVE, MUTE, SOLO, CUT — so the desk still reads with the colour taken away. The figures across the top are the extension's real ones. The meters are not: there is no audio here to meter.
04What it works out
The key and the tempo, read off the stems themselves.
The deck shows the detected key and tempo while the track plays. Neither is read from the mix, and that is the whole point: separating the audio first is what makes both estimates easy.
Key is taken from the other
stem. A drum kit is broadband, so it lifts all twelve chroma bins together and
flattens the profile toward uniform — measured at +3.5 dB RMS over
the harmonic content, enough to take a clean F♯ minor below the display gate. A
full-mix detector spends its life fighting the drums. This one does not have to;
the drums are in another lane.
Tempo is the exact mirror — taken from the drums stem, where broadband is exactly what you want. Three RMS band envelopes, a subtraction, and an autocorrelation over eight seconds of history.
Both taps sit upstream of the pitch shifter, so your transpose is added once, at one call site, rather than baked in twice. And speed is key-locked at every rate — 0.5× to 2.0× in 29 geometric steps, held to ±2 cents on every rung. Transpose is ±6 semitones, with the drums lane deliberately left unshifted.
When it is not sure, it shows nothing rather than a guess. Nothing downstream syncs to the tempo either — it is a readout, not a clock, so a wrong number costs you a wrong label and never a wrong render.
We have the drums already, so the front end is three RMS band envelopes and a subtraction — and that is not a shortcut we are getting away with, it is the one thing this product has that a full-mix detector does not.
- Key read from
- other stem
- Tempo read from
- drums stem
- Key resolution
- 16384-pt FFT, 2.69 Hz/bin
- Key decided at
- ~2 Hz, with hysteresis
- Tempo range
- 60 – 200 BPM
- Tempo history
- 8 s autocorrelation
- Speed
- 0.5× – 2.0×, 29 steps
- Speed moves pitch
- ±2 cents, every rung
- Transpose
- ±6 st, drums unshifted
- Written pitch
- concert · E♭ · B♭
- Tempo drives playback
- never — display only
The key can be read in a transposing instrument's own written pitch — concert, alto/bari in E♭, or tenor/soprano in B♭ — so a horn player reads the name they would actually finger. The engine only ever sends the concert tonic; the horn and your transpose are applied once, on the way to the screen.
A version with no delay would be a version that can see the future.
Latency is inherent, not a bug. About 3.4 s behind the picture at the default hop — measured, and shown in the deck as the exact figure rather than a claim.
06Claim one — what it can hear
It only ever hears what your own player renders.
Audio comes from chrome.tabCapture and nothing else. No stream-URL
resolution, no yt-dlp, no player-response scraping. If your player
isn't making a sound, neither is this.
It cannot save a file. There is no downloads
permission in the manifest, and an automated CI check asserts its continued
absence on every run. This is the line between an audio tool and a ripper, and
it is enforced by the build, not by a promise.
Export specifically is not coming back, because writing a separated copy of somebody's track to disk is the thing this project exists not to do.
- Audio source
- chrome.tabCapture
- Stream-URL resolution
- none
- yt-dlp / scraping
- none
- downloads permission
- absent
- Absence asserted by
- CI, every run
- Export feature
- not coming back
- Bytes uploaded
- 0
07Claim two — what it sends
It makes exactly one network request, ever.
The model weights, once, from a pinned and hashed host. After that: no telemetry, no analytics, no error reporting, no fonts, no update pings. A full session completes with every network interface disabled — that is an acceptance test, not an aspiration.
One host for the weights is a single point of failure, and the project says so in its own documentation rather than waiting for you to find out.
This page holds itself to the same rule. No CDN, no web fonts, no analytics, no third-party anything: one HTML file and its own images. A site for a product whose headline claim is one network request cannot itself phone home.
This is a real single point of failure and we would rather name it than have you discover it.
- Requests, lifetime
- 1
- What it is
- model weights, 109 MB
- Host
- pinned + hashed
- Telemetry
- none
- Analytics
- none
- Error reporting
- none
- Web fonts
- none
- Update pings
- none
- Offline session
- acceptance test
08Signal path
Everything between the tab
and your speakers.
Eleven stages, one clock. The model runs on the last 7.8 s of a rolling buffer and emits only the newest hop, which is why the delay is a property of the design rather than a defect in it.
- 00 YouTube tab
- 01 chrome.tabCapture
- 02 AudioContext @ 44100 Hz
- 03 Capture worklet → SharedArrayBuffer ring
- 04 Causal chunk plan
- 05 ONNX Runtime Web + WebGPU · htdemucs_6s
- 06 Weighted overlap-add
- 07 14-plane stem ring
- 08 Optional transpose
- 09 Playback worklet
- 10 Your speakers
offscreen document
The engine — the only context with getUserMedia, workers, and SharedArrayBuffer.
content script
Injects the deck and owns the page's <video>.
the deck
An extension-origin iframe. Four Manifest V3 contexts, none of them interchangeable.
The architecture is the argument rather than the promise.
Which means the next section is not marketing. It is the measurement table.
10Receipts
Measured, gated,
and written down.
Every figure below is produced by the verification suite, against a gate that was
set before the number was known. A suite that exits 0 while asserting nothing is not
green: silence is not a pass
.
| Measurement | Result | Gate / control |
|---|---|---|
| Sample-clock round trip, capture → model → output | 1000.00 Hz | one AudioContext at 44100 Hz, no JS resampling on the live path |
| Overlap-add reconstruction, identity model | −160.6 dB | −120 dB |
| COLA error, Float32 | 3.0e-8 | 1e-6 |
| Six-source synthetic separation null | −144.5 dB | worst-single-omission control: −6.3 dB |
| Pitch shifter bypass identity | −294.5 dB | −120 dB |
| Pitch shifter alias floor | −116.9 dB | −60 dB |
| Pitch shifter passband, 40 Hz – 19 kHz | ±0.02 dB | ±0.5 dB |
| Key-lock: speed must not move pitch | ±2 cents | across all 29 rungs |
| Transpose group delay | 3072 samples | 69.66 ms, exactly, on every lane at every setting |
| Verification gates | 19 gates · 1302 assertions | 17 gates (1160 assertions) run with no browser and no weights |
| Model | htdemucs_6s | 27.4 M params · 109 MB download |
| Latency | ~3.4 s | at the default hop — measured, and displayed in the deck |
| Network requests, lifetime | 1 | the weights, once |
Two numbers this project refuses to quote
Because a number you cannot stand behind is worse than no number at all.
- Any millisecond figure from the pitch bank. Three runs of identical code swung 69 %. That is the machine, not the code, so frame counts are quoted instead.
- Any repo-wide timing absolute. Same reason. There isn't one on this page either.
11Install
Three commands
and a folder.
There is no Chrome Web Store listing. You load it unpacked, from source you can read.
bash
$ git clone https://github.com/itziklerner-pag/stem-splitter-live
$ cd stem-splitter-live
$ bash tools/fetch-vendor.sh # ONNX Runtime WebGPU runtime, ~26 MB, not in git- Open
chrome://extensions. - Enable Developer mode.
- Choose Load unpacked and select the
extension/directory. - On first run a welcome tab offers the 109 MB model download. That is the one network request.
- Open a YouTube video and press CtrlShift9.
- Browser
- Chrome 128 or newer.
- Acceleration
- WebGPU strongly recommended. The WASM fallback works, but blocks for seconds at a time.
- Disk
- ~26 MB of runtime, fetched by the script; 109 MB of weights, fetched on first run.
12Keyboard
The map.
These are the extension's real bindings, and on this page they drive the demonstration deck in 03 · The surface. Press one now — the row lights up and the deck answers.
| Key | Does |
|---|---|
| 1–6 | Mute a stem — vocals, drums, bass, other, guitar, piano, left to right. |
| Shift+1–6 | Solo a stem. |
| Alt+1–6 | Reset that stem's fader to unity. |
| 0 | Unmute everything and clear the solo. |
| ? | The full list. |
| Ctrl+Shift+9 | Show or hide the deck. ⌃⇧9 on macOS |
13What's next
A roadmap, not a set of promises.
Nothing below has a date, and anything below can be cut. Two features already were — an offline export mode and a two-deck console were both built and both removed, and the architecture appendix records what went and why, so a re-proposal starts from that rather than from scratch.
-
01
Practice loops
Mark a section of a track and loop it, so a phrase can be drilled rather than scrubbed back to by hand. It composes with what is already here: loop four bars, drop the vocal, take it to 0.75× in the original key.
-
02
Notation from a stem
Generate readable notation from a separated stem, so the line you are learning can be read as well as heard — and read in your own written pitch, which the key readout already knows how to do.
-
03
A second deck
Two decks, for beat-matching and DJ use. This one has been built before and cut. Bringing it back is a real proposal rather than a wish, because the tempo detector and the key-locked speed control it would need are already shipping.
More beyond these. The place to argue for something, or against something here, is Discussions.
14Licence, honestly
Free, and staying free.
Two licences, because two different things are involved, and pretending otherwise would be the easy lie.
MIT
Take it, fork it, ship it. All of the capture, the ring buffers, the overlap-add, the pitch bank, the deck.
CC BY-NC 4.0
htdemucs_6s is Meta's, and non-commercial. They are not ours to license. The repo has never contained them; they are fetched at runtime from a pinned, hashed revision.
No paid tier
There is no paid tier and no plan for one, because a commercial product built on non-commercial weights would be a licence violation. That is the whole reason, stated plainly.
We could have quietly said 'MIT' and let people assume. Publishing a licence claim you cannot support is a worse problem than the constraint itself.