github.com/JonasGrunau/open_audio_analyzer
A free and open-source loudness and spectrum analyzer. Fourteen meter modules on a canvas you arrange — on the desktop, inside your DAW, and on a tablet beside the desk — measuring against ITU-R BS.1770-4 and EBU R 128, held to the reference cases on every push.
Measured from Apathy And Resignation Are The Sources Of Our Misfortune by Citizens Of The Empire, CC BY 3.0. The readings are the engine's own, recorded from that track — not a simulation.
Modules
Every module is one ModuleFrame plus a painter, reads the same
MeterSource, and repaints from the same clock. That is the whole reuse strategy: a
module that also owned its border and title treatment is a module that would drift from the
other thirteen. Independent tickers drift too — and two meters showing the same quantity that
disagree inside one frame is a correctness bug, not a cosmetic one.
Any single measurement, as a number.
Momentary and short-term loudness as bars, integrated loudness as a line.
Momentary, short-term and integrated loudness as three concentric arcs.
Sample peak and RMS, per channel, up to 7.1.
A needle, on the movement the engine models.
One measurement, watched, with the worst it has been latched.
The delivery decision, as a table.
Loudness against time: how the programme moved, and when it was over target.
How much of the programme was spent at each loudness.
Level against frequency, log-spaced, with a peak hold.
Frequency against time, with level as brightness.
The waveform itself, one lane per channel or both overlaid.
A goniometer: left against right, rotated so mono stands upright.
Where each frequency sits in the stereo image, accumulated over time.
Measurement
| Quantity | Definition |
|---|---|
| K-weighting | ITU-R BS.1770-4 stage-1 shelf plus stage-2 RLB, with the coefficients computed from the analog prototype at the actual sample rate — not from a hardcoded 48 kHz table. |
| Gating | EBU R128: 400 ms blocks at 75% overlap, absolute gate −70 LUFS, relative gate −10 LU. |
| Momentary / Short-term | 400 ms and 3 s windows. |
| LRA | Gated at −20 LU relative, 10th to 95th percentile, through a 0.01 LU-bin histogram that stores exact energy sums — O(1) per update, constant memory. |
| True peak | BS.1770-4 Annex 2: 4× oversampling through the specified 48-tap polyphase FIR, at every sample rate. |
| Spectrum | A 4096-point Hann window at a 1024-sample hop, zero-padded to a 16384-point transform and mapped onto 512 log-spaced bands with peak-per-bin, so narrow peaks survive. A full-scale sine reads 0.0 dBFS on a bin centre and within 0.3 dB off it. |
| Correlation | Running Pearson over a sliding window. |
| Crest | Sample peak minus RMS over the same block — the block’s own values, not the held peak and the smoothed RMS the meters draw, whose difference drifts on its own. Exactly 3.0103 dB for a sine, 0 for DC. |
| Clip | Longest run of consecutive samples at or above 0.999 since the reset, per channel. Latched, so a clip that lasted three samples is still visible when you look back. |
Dynamics
Decibel reports a dynamics figure called TrueDyn. It is proprietary and undocumented, so any claim to match it would be a guess presented as a measurement. Open Audio Analyzer does not implement it.
Instead it reports DR-S and DR-I, defined as
TruePeak − LUFS-S and TruePeakMax − LUFS-I, published in the
metrics document and reproducible from the definition by anybody who wants to check.
PSR and PLR are the same two quantities under the names in wider
use, and are offered as well — four names for two numbers, each computed once so the pair
cannot drift, and printed once in a report rather than twice under different headings.
Architecture
Measurements never cross an isolate boundary, never allocate per frame, and never rebuild a widget.
Capture
audio callback, realtime-safe
Copy input into a lock-free ring. No malloc, no locks, no syscalls.
Analysis
dedicated thread, high priority
Run the DSP graph. Publish results into a seqlock-protected snapshot.
Display
the UI thread
One FFI call per frame, then paint. The widget tree is never involved.
Once per frame the clock makes a single leaf FFI call — an atomic load, a memcpy, and a second atomic load, with no VM state transition. Then every painter reads views that were built once at startup over native memory that never moves. A frame costs one FFI call plus N paint calls, and allocates nothing.
The snapshot is a seqlock precisely because it is wait-free for the analysis thread: the reader retries, never the writer. A mutex would let a descheduled UI thread stall the thread that must never stall — and when that thread falls behind, the ring overruns and signal is lost for good.
Plugin, tablet and CLI
A VST3 and an Audio Unit that draw nothing. Insert one on a track, a bus or the master, and the desktop app meters what your DAW is playing — through the same engine, on the same canvas, with the same painters as a live input. The plugin measures and streams; the app displays. That split is not a compromise around a UI toolkit, it is what stops there being two implementations of every meter drifting apart.
Flip PUBLISH in the desktop’s status bar and a tablet mirrors the canvas over Wi-Fi. It finds the host by itself, or reads a pairing code off the desktop’s screen with its camera, or takes an address typed by hand — three routes, because the first one is what a venue’s Wi-Fi blocks. A preset stores grid cells rather than fractions of a window, so the same layout opens on a 32-inch monitor and an 11-inch tablet with nobody writing responsive code.
oaa decodes a file and pushes the blocks through the same analysis a
capture device drives — there is no second DSP path, and a test asserts that an offline
reading and a live reading of the same audio are identical rather than merely close.
Nothing resamples or remixes: a converter in front of a measurement changes the
measurement.
$ oaa master.wav # human-readable report
$ oaa --target streaming-14 master.wav # …and a delivery verdict
$ oaa --format json --timeline master.wav # every measurement, for scripts
$ oaa --format csv -o loudness.csv master.wav # the loudness timeline
$ oaa --list-targets # what you can measure against
# fail a release build on a master that is 2 LU too loud,
# instead of shipping it
$ oaa --target streaming-14 master.wav || exit 1Configuration
Settings, presets, delivery targets and skins are plain JSON in a documented directory —
open them, edit them, copy them between machines, keep them in version control. A skin
names as many of thirteen semantic colour roles as it likes and inherits the rest,
so changing one colour is a three-line file. The roles are semantic rather than literal —
over is “the colour that means over a limit”, used for nothing else — which is
what lets a skin apply to a module written after it.
There is an editor for them too. Every role gets a colour picker, the canvas repaints as you drag — and so does any tablet mirroring the session — and each role shows its contrast ratio against the surface it has to be read on, so a palette that puts a meter track 1.1:1 against its own panel says so before you ship it. The two skins that come with Open Audio Analyzer are fixed; copying one takes a press.
A delivery target whose id matches one that ships replaces it
everywhere, including in presets that already name it. So if you disagree with our
reading of a published spec, your number wins, and it reaches the CLI’s exit code and not
only the window. Delete the file and the original comes back.
{
"id": "house-standard",
"name": "House standard",
"lufs_target": -12.0,
"lufs_tolerance": 0.5,
"true_peak_max": -1.0,
"lra_max": 14.0,
"vu_reference": -18.0
}Known gaps
The list in the README is honest rather than short, and it is kept there rather than here so it stays next to the code. Six of the twelve are below.
The protocol frame and the engine’s half are both built; what is left is the part you would see — a mode menu and a region editor. Continuous is what every module measures today, and it is the mode a producer with no playhead can honour anyway.
WAV, AIFF, RF64, Wave64, FLAC and MP3 cover what a master is delivered as, which is what a delivery check is for. No measurement is silently wrong in the meantime: an unsupported file is refused rather than half-read.
There is no region selection and no seeking during analysis, because an integrated loudness taken over a file that was seeked through is a measurement of a programme nobody played.
Anyone who can reach the port can read the measurements and the layout — not the audio, which never goes on the wire. That is why publishing is off until you switch it on, and why the link is one-directional. Do not switch it on at a venue whose Wi-Fi you do not control.
An Android emulator sits behind NAT that does not carry the LAN’s multicast, an iOS simulator is exempt from the restriction the device enforces, and a test host is refused the local network for reasons that have nothing to do with the code. Both paths are checked by hand on hardware, and typing an address is supported everywhere and always will be.
Signing needs certificates that are not ours to commit, so a release built from a fork is unsigned and every packaging script says which of its three states it was in — because “signed” and “a user can double-click it” are not the same thing.
Installing
| Platform | Artefact | Plugin | Note |
|---|---|---|---|
| macOS 14.2+ | .pkg | VST3 + AU | Universal — Apple silicon and Intel |
| Windows 10 1809+ | .exe | VST3 | Uninstaller in Installed apps |
| Linux | .tar.gz | VST3 | install.sh, no root |
| Linux | .AppImage | — | One file, no root, GTK from the host |
| Linux | .flatpak | — | Sandboxed, carries its own runtime |
| Any | oaa CLI | — | The analyser. No Flutter runtime. |
The first three install the plugin as well, and the checkbox starts ticked. Metering your own system’s output needs nothing installed on Windows, on macOS 14.2 and later, or on Linux — no driver on any platform, and on macOS the audio still reaches your speakers while it is measured.