You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
A USB cable needs nothing switched on. An iPad plugged into a publishing
desktop now appears under Over USB with no developer mode, reached
through the service Finder syncs it with — built into macOS, usbmuxd on
Linux, the Apple Devices app on Windows. An Android tablet plugged into a
publishing Mac or Linux desktop is asked to become its USB accessory, Android
offers to open Open Audio Analyzer, and it appears in the same place with no
USB debugging and no adb. A device is asked once each time it is plugged
in, and never while PUBLISH is off; while it is the desktop's accessory its
files are not offered to the desktop. On Windows an Android tablet still needs
USB debugging and adb. (#5)
A tablet on a USB cable is a display. An Android tablet plugged into the
desktop now shows up in ATTACH under Over USB, marked with a plug where a
network host has the broadcast mark, and attaching to it bypasses the room's
Wi-Fi. While PUBLISH is on, the desktop forwards its display port to every
Android device adb can see (adb reverse), so it needs adb on the desktop
and USB debugging on the tablet. A host found over USB tethering is listed
in the same section. An iPad has no equivalent the desktop can drive, so this
is Android only. Attaching a machine to itself is still refused; what is
refused now is this running copy, not every address on this machine. (#5)
ATTACH remembers. Every host this device has been a display for is listed
under Recent, newest first, up to five, with its name and address, and
each one can be forgotten. A host is remembered only once it has answered, so
a mistyped address never comes back. In a room that blocks discovery that is
the difference between typing an IP address at every session and tapping it
once. (#5)
A display keeps the screen on. While a host is attached, an Android or
iPad display no longer goes to sleep, and the system's own sleep setting
applies again the moment it disconnects. On by default; OPTIONS on the display's bar
turns it off. (#5)
The display has options of its own: OPTIONS on its bar opens This display, with the refresh rate — the one the canvas on this device uses
— and the screen setting above. A tablet that falls behind a heavy layout can
now be told to redraw at 30 fps; before this it ran at 60 whatever it was
set to. (#5)
A single tab can be saved, loaded and put back. A tab's menu has Revert to
saved, which returns it to the preset as it was last opened or saved in one
undoable step, and Export…, which writes it as a preset with one tab. File › Import tabs… adds tabs from any preset file beside the ones already there,
asking which when there are several, and Export this tab… exports the one
showing. A tab file is an ordinary preset, so File › Open reads it too.
Undo was not enough: it is shared by every tab and forgotten at the next
launch. (#5)
An AAX build for Pro Tools. The plugin now builds a fourth format on macOS
and Windows, universal on macOS and targeting 14.2 like the other bundles, and
it ships in the oaa-plugin-<platform> release archive. It is not signed
for Pro Tools: an AAX bundle needs a signature from PACE's wraptool made
against an Avid developer account holding a signing certificate, which is a
different thing from the code signature every bundle already carries and which
this project does not have yet. Unsigned, it loads in a Developer build of Pro
Tools and in Avid's developer tools and nowhere else, and a released Pro Tools
refuses it by simply not listing it. That is why no installer carries it — a
ticked checkbox that installs something no DAW will show you is worse than no
checkbox. The bundle passes Avid's own AAX Validator: it describes itself,
its Describe function validates, every effect instantiates and de-instantiates
under every host context, and it survives a thousand load/unload cycles.
The website has a page at /alternatives/decibel for readers arriving from a
search for a free alternative to Decibel by Process.Audio. It states what the
two share, the four differences there is published evidence for, and the
repository's Known gaps rather than a softened version of them. The claim that
this project reimplements Decibel's ideas has been in README.md since the
first commit and was on no page of the site.
The website has an Impressum at /impressum — the provider identification
§ 5 DDG requires of a German site: name, postal address, telephone, email, and
who is responsible for the content under § 18 Abs. 2 MStV. German first,
because that is the word a German reader looks for in a footer, with an
English translation under it. Both footers link it as "Legal Notice
(Impressum)" — the English for a site written in English, and the German word
because that is the one § 5's recognisability is judged by — and it is in the
sitemap.
Open Audio Analyzer is on the App
Store, free, for iPad — and the
front page's App Store badge is a link to it. The badge stood there dimmed
and unlinked through the wait for review, which said "not yet" to nobody who
did not already know it was waiting; that state and the opacity rule behind
it are gone with the wait. The store carries the build Apple has passed
rather than the version tagged last, so the listing can sit behind the
desktop downloads — TestFlight still takes every tagged release first, and
Installing says which is which.
--tab=<n> opens the application on a tab, counted the way the tab strip
counts. It exists for the screenshot scripts, which no longer press a key at
the window — see Internal — and it works on every platform.
The website and the documentation name the DAWs the plugin has been tested
in — Ableton Live, Logic Pro, Cubase, FL Studio and Studio One — and say what
it costs in so many words: free, with no paid tier, no trial and no account.
They also state the channel ceiling, eight, mono through 7.1. "Free and
open-source" was the only price on the site, and it reads as a licence to
anyone looking for a price.
⚡ Changed
The two dynamics readings are printed as PSR and PLR again, the
names the AES gives the arithmetic and the ones every other meter prints —
and Settings › Dynamics switches every label to the specification's own ODR-S / ODR-I: on each module, in the metric picker, in the target
editor, in the report panel, card and text export, and on a paired tablet,
which is sent the choice beside the skin and the target. 0.15.0 renamed the
pair to ODR-S / ODR-I alone, and a person who typed PSR into the metric
picker found nothing and concluded the meter did not measure it. Nothing
measured has changed and no re-measure is warranted: the arithmetic, the
ids in a preset, the odr_i and odr_s_min keys of a JSON report and the odr_i_min / odr_s_min floors of a delivery target are all as they were.
The oaa CLI takes the same choice as --names psr or --names odr,
which spells the text report and --list-targets; it reads no settings
file and still does not.
🐛 Fixed
An Android tablet is called Android tablet where it names itself — as a
publishing host, and in a desktop's list of tablets on a cable — rather than localhost, which is all Android tells an application its name is.
Save as, Export this tab and the report exports work on an Android tablet
and on an iPad. They opened nothing and saved nothing on either, because file_selector has no save dialog on Android or iOS; they now open the
system's own document picker. A preset opened on a tablet is now the file
that was picked rather than a copy of it, so Save writes back to that file —
it used to write to the copy, succeed, and leave the file untouched — and the
file is remembered across launches like on every other platform.
A display that falls behind is no longer drawn seconds in the past. A
tablet slower than the frames it was sent — a heavy layout on an Android
tablet was the report — worked through a backlog in order, so the delay grew
and nothing on screen said so. The host skipped a display only while its last
write was still leaving the machine, and every network and kernel buffer
between the two filled before that happened; in a test, a display that stopped
reading was queued 62 of 120 frames. A display now tells the host each time it
has read a snapshot, and the host sends it at most a tenth of a second ahead
of that, dropping the rest. A display from an earlier release still works, and
is fed as before. (#5)
The privacy policy was 128 px wider than a phone screen. The address to
report a privacy question to is printed as a link whose text is the URL, and
a 57-character word has no break opportunity in it — so on a 390 px phone it
laid out at its own width and pushed the document sideways. Nothing scrolled,
because the page cannot scroll sideways, so what a reader saw instead was a
page 128 px narrower than the window from the header down, with every heading
and every paragraph short of the right edge. Long links now break.
Every page carrying the one-line footer was 37 px wider than a 360 px
phone, for the same reason and on a smaller scale: the row of five links in
it did not wrap. Small enough to look like nothing, and on every documentation
page, which is the combination that makes a bug read as the design.
npm run check reported every page clean while three of them overflowed.
It opened each page in an emulated mobile viewport, and mobile emulation
scales a too-wide document down until it fits — so the width the page reported
was the width of its own content and the check compared a quantity with
itself. It now measures at the width it asked for, and fails outright if that
is not the width the page got.
The oscilloscope on the website draws a continuous waveform. It drew a
comb — half a beat of trace, a gap the same width, over and over — on both
the front page's still and the live analyzer behind it. The recording is not
at fault and carries no audio: the browser replays the engine's readings and
reads the waveform out of the track it is already playing, and it was handing
the module one 1,024-sample block for every 2,205 samples the clock said had
gone past. The module is right to draw the shortfall as blank columns —
audio that was measured and lost is not a straight line — so the fix is the
window, which is now the four blocks the interface allows for and covers any
cadence down to eleven frames a second. The waveform also no longer runs nine
per cent fast on a machine whose audio output is not at the track's sample
rate: both web targets now decode at the recording's rate, and where a
browser refuses that, the window is stepped into the reading's time base
instead. Nothing the application, the CLI or the plugin measures is affected
— this is the website's replay source alone, and it now has a test.
The website's Modules section told readers to drag a meter onto the
canvas, which is not a gesture Open Audio Analyzer has or has ever had:
there is no palette to drag a meter from, and dragging on the canvas moves a
module that is already placed. It now names the ones that exist — right click
the canvas, or long press it on a tablet — which is what README.md has said
since the first release, so the first instruction on the page can be followed.
The credit under the front page's live analyzer also says what that canvas is:
one measured pass of the track, replayed on a fixed eight, and not the canvas
the section two screens below is about. (#4)
The plugin told every DAW it was 0.11.0. The VST3, the Audio Unit and the
AAX have all declared that version since 0.12.0 — in each bundle's version
field and in the VST3's moduleinfo.json — while the installer that carried
them put the real version on the application, so a host listed a plugin four
releases behind the program it streams to. The bundles were built from the
right commit every time and nothing they measure was affected; the number was
a literal in plugin/CMakeLists.txt that a release commit was expected to
bump, and after 0.11.0 nothing did. It is now read from pubspec.yaml, and
the macOS package refuses to build if a plugin bundle's version is not the one
on the box — which is also what catches a stale build tree. A DAW that still
shows 0.11.0 after this release has cached the old bundle, or is loading a
copy from ~/Library/Audio/Plug-Ins, which no installer here writes to or can
remove.
GNOME Software and KDE Discover offered 0.11.0. The flatpak's AppStream
metadata lists a row per release and its newest row is the version a software
centre prints; it stopped being written after 0.11.0, so four releases
installed while the store beside them described the one before them. Every
release through 0.15.0 is listed now.
The AppImage's desktop entry passes the validator on Ubuntu 22.04. That
version predates the SingleMainWindow key and rejects it, and it is the one
AppImageHub's catalog test runs, so 0.15.0 was refused a listing on that one
line. The AppImage now carries GNOME's older X-GNOME-SingleWindow, which
gnome-shell reads the same way; the flatpak keeps the standard key.
🚧 Internal
docs/WIRE.md defines 0x0007 RECEIVED, the one frame a display sends on the
display port: empty, once per snapshot, read by the host only to decide what
not to send. Added under protocol version 5 without moving it, the way 0x0006 was; a host that predates it never reads the socket, and a display
that predates it is judged by the flush alone. The display port still carries
no control of any kind. (#5)
A fourth Android channel and a third iOS one, OaaKeepAwake, over com.openaudioanalyzer.oaa/keep_awake: FLAG_KEEP_SCREEN_ON on the
activity's window, and isIdleTimerDisabled. (#5)
docs/WIRE.md defines the USB carriage: how a tablet's relay carries
each display connection as a channel of one cable, on loopback 47823 and
47824. It changes no table of the display port, and inside a channel the
bytes are the display port's. (#5)
packages/oaa_usb, a fifth local package: the Android Open Accessory
protocol over a vendored libusb 1.0.30 (LGPL-2.1), built on macOS and Linux
and tested by the engine job on all three. OaaAccessory.kt pumps the
accessory into the tablet's relay and is not a channel. (#5)
A fifth Android channel and a fourth iOS one, OaaDocuments, over com.openaudioanalyzer.oaa/documents: the system's create and open pickers,
and reads and writes through the ContentResolver on Android and a
security-scoped bookmark on iOS. A picked document travels through the app as <handle>#<display name>, routed by ConfigStore, because a handle need not
contain the file's name.
robots.txt names the AI crawlers rather than leaving them to User-agent: *
— OpenAI's three agents, Anthropic's three, Perplexity's two, Common Crawl,
Meta's two, Mistral, DuckDuckGo, Amazon, ByteDance, and the Google-Extended
and Applebot-Extended control tokens, all allowed everything. It grants
nothing that was not already granted; what it fixes is that a crawler reads
exactly one group and never merges it with *, so a Disallow: added to * in future would have silently exempted every one of them. The file records
that it is only half the policy: Cloudflare can refuse an AI crawler at the
edge before the Worker is reached, so the check is a request wearing the
agent's own name, and the curl for it is in the file.
The screenshot scripts take nothing from the person at the machine. packaging/signal_path.sh, packaging/app_window_shots.sh and packaging/ios/screenshots.sh used to bring the application and the
Simulator to the front and post keys at them — the tab digit and the ⌘→ that
turns the iPad — so a run stole the focus from whoever was working, and was
interrupted for it. They post nothing now: both applications open in the
background with open -g, the tab is the --tab launch option, and the
device's orientation is checked and never set, because every way of turning
it from a script works only while the Simulator is frontmost. A booted device
is kept as it is, and a portrait one is refused with the menu path to use. The
desktop window no longer has to be uncovered either: macOS reports a window
visible while any part of it shows, which is the signal Flutter pauses on,
and the shutter reads the window's own image — so a corner showing from
behind a browser is enough, and the run says how much showed.
signal_path.sh shot a portrait iPad. It wrote SimulatorWindowOrientation
and a rotation angle into the Simulator's preferences and then pressed ⌘→, on
the theory that the pair shapes the window and the key turns the device;
present, the pair pins the device so that nothing turns it, and the press
went unchecked. The keys are removed before launch, as the iPad script does.
ODR is at 1.1. § 6.4 permits the AES names beside the specification's
own and asks an implementation to say which it prints; § 2 cites the current
AES document, TD1008, and the 2017 convention paper that proposed PSR, in
place of the superseded TD1004; and § 8's PSR / PLR row no longer claims the
AES leaves the peak open — it does say true peak. What it leaves open, and
the specification pins, is the channel, the window's alignment, the gate,
the statistic and the display. No operand moved and every conformance case
reads as before.
The wire has a 0x0006 DYNAMICS_NAMING frame, host to client, carrying the
names as settings.json spells them. Added under protocol version 5 without
moving it: it changes no table, and a display that predates it skips it by
length and prints the default, which is what a host that predates the
setting prints too. docs/WIRE.md says so.
The website has photographs of the application's own window — the Loudness
tab and the Spectrum tab, chrome included — taken by packaging/app_window_shots.sh and encoded by npm run window. Nothing on
the site showed the tab strip, the file name, the menu row or the status bar:
the hero still is a Flutter web build with no window around it, the module
thumbnails are one module each, and the signal-path plate shows the window
mid-publish because that is what its section is about. Each is a session of
the fake DAW playing a real track through the real VST3, so every reading in
them is one the engine took.
packaging/ios/screenshots.sh no longer touches the pointer. The App
Store set is three pictures now — the Loudness tab, the Spectrum tab, and the
Loudness tab in the Daylight skin — and nothing with screen coordinates is
posted to take them: the tab is the application's own digit binding sent as a
key, the skin is written into the app's settings.json before the launch that
photographs it, and the rotation is a key whose effect is read back off the
Simulator's window bounds and retried until the device is on its side. The
old five posted mouse clicks at coordinates read off a finished screenshot,
which drifted twice and pressed the wrong controls without failing; the
delivery-target, module-library and remote-display pictures are gone with
them.
Every photograph is committed.packaging/ios/screenshots/ holds the App
Store set, packaging/android/screenshots/ the Play Store set, and packaging/screenshots/ the website's four plates — the signal-path pair and
the two window shots — and the four scripts write there now rather than into build/packaging/, which a clean deletes. Nothing in CI can take these
pictures — they need a fake DAW, a simulator or an emulator, and a person's
grants — so unlike an installer they could not be rebuilt from nothing.
The Windows engine job no longer fails about one run in four. The reclaiming orphans tests called oaa_engine_reset_all, which destroys every
live engine in the process — and dart test runs its suites as isolates
inside one process, four of which create engines. So the reset reached into
whichever sibling suite was running and freed engines it still held, and the
job died with an access violation inside the engine library. The three cases
now run as a program in a process of their own. Nothing the engine does was
wrong and no measurement changes; the test was.
website/tools/oaa_replay has a test suite, run by the checks job. That
package is not a workspace member, so flutter analyze never reached it and
the only thing that read it at all was dart format.