Long-form pieces on how this stuff actually works — the mechanisms, the failure modes, and the measurements behind them. Written because the explanations we wanted did not exist.
| Piece | What it covers |
|---|---|
| Make a Karaoke Track from Any Song |
Turn any song into a karaoke backing track in the browser, with no upload. The process, what the result actually sounds like, and how to tell in advance whether your song will separate cleanly. |
| Three audio tools in one browser tab |
A walkthrough video of all three tools, what each one downloads before it can start — 52 MB, 16 MB and 64 MB — and the four-step network-tab check that settles whether your file is being uploaded. |
| A 130 MB WASM model on Cloudflare Pages |
Cross-origin isolation, a 25 MiB per-file wall, and a host that reports the opposite of the truth when you ask it about compression. Four failures from shipping a pthreads WASM model to a static host, three of which produce no usable error message. |
| How to remove vocals in Audacity |
Two ways to do it, why the instrumental comes back hollow, and a 30-second test for whether your file will work at all. Includes a measurement: the peak envelope drops by about 40 per cent, and only part of that was the voice. |
| Honest comparisons: LALAL.AI, Moises, vocalremover.org |
Where each of them beats us, said plainly: metered minutes buy more stem types, an account buys a synced library, and a server buys full-length files. Three separate write-ups behind one page, each with the case for using the other tool. |
What tends to end up here
Not tool round-ups and not release notes. Head-to-head comparisons do appear, but only the kind that names what the other tool does better. The two subjects that keep producing material are the signal processing — what phase cancellation can and cannot reach, why separation artefacts land where they do — and the delivery problem of running a machine-learning model inside a browser tab at all.
Both are written the same way: state the mechanism, then measure it. Where a number appears it was measured on the actual artefact, not estimated.
Frequently asked questions
Can I reuse the diagrams?
Yes, with attribution and a link back to the piece they came from. The diagrams are drawn as inline SVG rather than exported images, so they stay legible at any size and their labels remain selectable text.
Do the tools appear in these articles?
Where they are the honest answer to the question being asked, and not otherwise. The Audacity piece spends most of its length on Audacity, because that is what it is about.