WavRead
Device requirements
What WavRead needs, what it runs well on, and where its ceiling actually is. Every number here was measured on a real machine rather than estimated — how, and on what, is at the bottom of the page.
Minimum
| Mac | Apple Silicon — M1 or later. WavRead ships as an arm64 binary only, so Intel Macs cannot run it, with or without Rosetta. |
|---|---|
| macOS | macOS 14 Sonoma or later. That floor comes from the Python runtime and the audio libraries built into the app; on macOS 13 or earlier it will not launch. |
| Windows | Windows 10 or 11, 64-bit (x64; Windows 10 version 1809 or later). WavRead draws its window with the Microsoft Edge WebView2 Runtime, which Windows 11 and an up-to-date Windows 10 already include; the installer checks for it and says so if it is missing. No administrator password: it installs for your account only. The Windows build is 536 MB installed, from a 137 MB installer, and is not yet code-signed — what Windows shows the first time. |
| Memory | 8 GB. Enough for songs of ordinary length — see the ceiling below, because memory is what actually limits WavRead, and it grows with the length of the track. |
| Disk | 2 GB free to install and work in. The app is 345 MB installed, from a 131 MB download. |
| Network | None. Analysis never needs it. At most WavRead makes five kinds of request: the once-a-day update check you can switch off, the update you press the button for, a one-time 460 MB model download if you transcribe lyrics, a problem report if you turn reports on, and device linking if you use it. WavRead ships no separation model and never fetches one. Privacy lists every one, field by field. |
Recommended
| Memory | 16 GB or more. It buys you long tracks without the machine paging — memory is what limits WavRead, and it grows with the length of the track. |
|---|---|
| Disk | An SSD with room to spare. Leave room for your exported mix, supplied stems, and saved analyses. You can choose an external drive for your WavRead library. |
| CPU | Any Apple Silicon chip, or any 64-bit Windows PC from the last several years. Analysis is the heavy step, and it is arithmetic on the samples rather than inference. |
Measured performance
One real 3-minute song, and the same song repeated to over 12 minutes, on the machine described at the bottom:
| What was run | Track length | Time | Peak memory | Written to disk |
|---|---|---|---|---|
| Analysis only | 3:06 | 5.1 s | 2.1 GB | 16 KB |
| Analysis only | 12:22 | 16.5 s | 4.4 GB | 32 KB |
Analysis runs roughly 36 to 45 times faster than realtime, and it is a one-off cost — the documents are written once and read from disk after that. Lyrics and melody are separate optional passes and are slower; they are transcription rather than measurement.
The ceiling: memory, not speed
WavRead holds the whole track in memory to measure it, so peak memory grows with length — measured at about 0.25 GB per additional minute of audio, on top of a 1.3 GB base.
| Track length | Expected peak memory | Comfortable on |
|---|---|---|
| 3 minutes | ~2.1 GB measured | 8 GB |
| 5 minutes | ~2.6 GB | 8 GB |
| 10 minutes | ~3.8 GB | 8 GB |
| 12 minutes | ~4.4 GB measured | 16 GB |
| 30 minutes | ~8.8 GB (extrapolated) | 16 GB or more |
The 3- and 12-minute rows were measured; the rest follow the straight line between them and are marked as such. macOS will page to disk rather than fail outright, so an 8 GB Mac given a 30-minute file gets slow, not broken — but a DJ set or a podcast episode is the case to watch.
Upper limits — what it will accept
| Sample rates | 44.1, 48, 96 and 192 kHz all tested and analyzed correctly. True-peak oversampling adapts to the rate rather than assuming 48 kHz. |
|---|---|
| Bit depth | 16- and 24-bit PCM, and 32-bit float. |
| Channels | Mono, stereo, and more — a 6-channel file analyses fine. Peak and clipping are measured on the real channels, never on a mono fold-down. |
| File formats | WAV, MP3, AIFF, FLAC, M4A, OGG. |
| Track length | No hard limit. Memory is the practical ceiling — see the table above. |
| Short files | Down to half a second. Measurements needing a longer window are reported as unavailable rather than guessed. |
| Stems per song | No fixed cap. Available memory and the size of the supplied files determine the practical limit. |
What it does not need
No internet
Analysis runs entirely offline. Nothing is uploaded, ever.
No app account
The desktop app has no sign-in, licence server, or telemetry. Your WavRead account lives on this website, where the downloads are.
No dedicated GPU
Every measurement is arithmetic on the samples, on the processor you already have.
No admin rights
No administrator password on either platform. On a Mac, drag it to Applications and open it; macOS asks once whether to open a program downloaded from the Internet, the ordinary confirmation, which also reports that Apple checked the app and found nothing malicious. On Windows the installer installs for your account only. The install guide shows exactly what appears on each.
Not supported
- Intel Macs — the build is arm64 only.
- macOS 13 Ventura and earlier.
- 32-bit Windows, and Windows 10 before version 1809.
- Linux — there is no build for it today.
- iPad and iPhone.
How these were measured: the Mac timing, memory and disk
rows are from an Apple M4 with 24 GB of memory running macOS 26.5, using
WavRead 1.4.47, and have not been re-measured since; later builds dropped a
separation backend and got smaller, not slower. Time and peak memory come from
/usr/bin/time -l around a real analysis; disk figures are the bytes
actually written. The macOS floor was read from the load commands of all 380
binaries inside the app bundle, taking the highest. Your own times will differ
with your chip and what else is running — the shape of the numbers is what
matters, not the last digit. The Windows sizes are the bytes of the 1.5.6
installer and of its installed folder on Windows 11 x64; the Windows floor is the
installer's own minimum version.