Qudelix for macOS

A native macOS menu bar app for configuring the Qudelix 5K DAC/amp, over USB or Bluetooth.

Download Downloads macOS 14+ Universal MIT

Unofficial, and not affiliated with or endorsed by Qudelix, Inc.

Qudelix ship a browser-based configuration app. On macOS it is unreliable — Chrome's WebHID goes through the same system call that, if a report is framed even one byte wrong, makes the 5K stop responding and drop off the USB bus. This app talks to the device directly through IOKit instead, and gets the framing right.

• Live device status — battery, charging, firmware, sample rate, input source
• Battery in the menu bar — a bolt while charging, the level once it runs low, and notifications at 20% and 10% (see below)
• Volume control with mute
• Sample rate — set the rate macOS runs the 5K at, change which rates the device offers, or let the app match the rate to what you're playing (see below)
• 10-band parametric EQ editor: filter type, frequency, gain, Q, plus pre-gain
• 20-band mode — switch between 10 and 20 bands from the EQ pane; the app also follows a switch made anywhere else
• Live response curve showing the combined filter shape, and — after an import the device could not take exactly — the requested shape behind it, so you can see where the two part company
• Curve editing — drag a band's dot to shape it, click one to open an inspector for its filter type, gain and Q under the graph, double-click to clear it; undo and redo throughout
• Auto pre-gain — works out how much the boosted bands need pulling back and offers the number, rather than moving it for you
• Per-band mute for an instant A/B of one band; the gain is kept
• 20 preset slots, which you can name from the app — the name is stored on the 5K, so other software sees it too
• Preset library on the Mac, beside the twenty slots on the device — as many curves as you like, global or bound to one output, applied through the same gated path as an import
• Headphones field — name the pair on the end of the 5K, and when the AutoEq project has measured that model a fitted correction is offered once
• Per-app EQ — give a single app a curve of its own from the library, applied on the Mac before the audio reaches the 5K, so a podcast player can run a speech-shaped preset while music keeps the device's curve (see below)
• Profiles — pair an output device with a preset and be offered the switch when that output becomes active
• Preset import from a file, from the clipboard — paste the filter list as published sites print it — from a file dropped onto the menu bar icon, or from the AutoEq database (6,000+ headphones), fitted live to a target curve of your choosing with bass and tilt adjustment
• AI preset studio — design a preset for the headphones you named, with an AI provider of your choosing on your own API key. It researches the model once and caches what it learned, then designs one of fifteen kinds of preset and shows it as a draft you can audition on the 5K, keep in the library, or write into a device slot. Folded away by default and silent until you press Generate; the key lives in the Keychain and nowhere else. Correction needs no provider at all when the AutoEq project has measured your model. See Privacy for exactly what is contacted and what is sent
• Device settings — channel trim, volume limit and the DAC reconstruction filter, all stored on the 5K itself
• Export your EQ in the standard parametric format
• Update check in the About panel — on request only, never in the background
• Tune — find the EQ you actually prefer, by ear (see below)
• Stage — a soundstage for headphones: width, crossfeed by band, balance, dialogue lift and room, with a true-peak limiter, loudness compensation, dynamic bass and convolution with an impulse response of your own, all applied on the Mac (see below)
• Level — live output level, an estimate of the level at the ear, a 14-day listening history, and a signal path inspector showing what is altering the audio and what is passing it through
• Call and microphone awareness — the Mac-side engine steps aside while the headset is on a call, and a guard tells you when an app grabs the headset's microphone and drops the Bluetooth link to the voice codec, or puts the default input back for you
• Stream quality detection — measures whether what's playing looks lossy or lossless (any player: it doesn't ask apps, it analyzes the audio), and can auto-match the USB rate: lossless → 44.1 kHz bit-perfect, lossy → your chosen rate. It is on by default, and listening is how it works: see The audio tap. Both parts can be switched off. When it has measured a cutoff, an import can also be told to stop correcting there, rather than spending filters and headroom on frequencies the source has already discarded
• Diagnostics panel logging every packet exchanged with the device

Works over USB or Bluetooth. USB is used whenever the 5K is plugged in; otherwise the app controls the device over Bluetooth LE. Either way it only speaks to the 5K's control interface, so playback is unaffected. Two features do alter the audio, and only while you switch them on: with the Stage running, or an app assigned under Per-app EQ, the Mac's audio is processed on its way to the output device (the 5K's own EQ still runs on the device, untouched). Two others listen to the audio without altering it, and one of those is on by default — all four are described under The audio tap.

Download the .dmg from Releases, drag the app to Applications, and launch it.

macOS will refuse to open it the first time — the app is signed only ad-hoc, not with a paid Apple Developer ID. To allow it, open System Settings → Privacy & Security, scroll to Security, and click Open Anyway. (The old right-click → Open trick no longer works on current macOS.)

If you would rather not trust a binary, build it yourself — see below.

Requires macOS 14 or later. Universal binary, Apple Silicon and Intel.

Because the app is signed ad-hoc, macOS cannot tell you who built it, and the "Open Anyway" click above is pure trust. The checksum published with every release is what narrows that gap. Before opening the DMG:

shasum -a 256 ~/Downloads/Qudelix-1.4.0.dmg

Compare the result against the SHA-256 in the latest release notes. Or, if you also downloaded the .dmg.sha256 file, let shasum do the comparison:

cd ~/Downloads && shasum -a 256 -c Qudelix-1.4.0.dmg.sha256

That should print OK. If the hashes differ, or the check fails, do not open the file.

Be clear about what this does and does not buy you: it catches a corrupted download or an asset replaced after publication, and it lets you confirm two people downloaded the same bytes. It is not a substitute for notarization, and it assumes the release page you read the hash from is itself genuine. Building from source sidesteps all of it.

The original Qudelix 5K on firmware 3.x, in either 10-band or 20-band EQ mode.

The app identifies the device during its connection handshake, and if it finds something it does not implement it says so and writes nothing, rather than silently doing the wrong thing:

EQ is personal, and reading a frequency-response graph tells you very little about what you will enjoy. The Tune tab offers four ways to settle it by ear. Every pair you hear is matched for loudness on its magnitude response across the range where music lives, so the louder option never wins by being louder.

Tune

Blind A/B. Play music you know well, and the app presents two EQ settings; you pick whichever sounds better. It narrows down from there, over roughly twenty comparisons, and ends with a curve you can keep or save to a preset.

1. Start playing music at a comfortable volume, and leave the volume alone.
2. Open Tune → Compare → Start.
3. Use Switch to flip between A and B as often as you like, then Prefer A or Prefer B. If they genuinely sound the same, say so — that answer is used, not discarded.
4. At the end, Keep the result or Save to slot.... Discard restores exactly what you had.

Both options are always matched for loudness and never labelled, so you cannot simply prefer the louder one — which is what happens in most casual A/B tests. Some pairs are deliberately identical, as a check on how reliable the session was. If you named a winner on most of those, the session says so and refuses to write a result.

A shorter route to a curve: bass, then presence, then overall tilt, one axis at a time, three rounds each, halving the step. Three controls settle in twelve comparisons where the full Compare takes about twenty, and you are never asked about tilt while the bass is still moving underneath it.

Each axis is fitted to your live band centres, so what you hear during the session is exactly the curve Keep will write — not an approximation of it. When the pre-gain headroom cannot carry the full ranges with matched levels, the ranges shrink until it can, and the intro says what fraction is being explored. Three axes landing within what anyone can reliably hear is reported as a real answer: leave it alone.

Does the EQ you already have survive not knowing which side is which? Five trials, sides randomised, your curve against a flat one — a flat curve, never the enable switch, because switching the EQ off takes the pre-gain with it and turns the whole thing into a test of which side is louder. Four of five is the line for saying which way you leant; anything less is reported as the difference not surviving blinding, with all three counts shown. It writes nothing whatever the answer.

Plays faint tones at ten frequencies and finds the quietest level you can hear at each, then compares that with typical hearing.

1. Set a comfortable volume, and sit somewhere quiet.
2. Open Tune → Tones → Start.
3. Press I hear it (or the space bar) the moment you hear anything, however faint. Roughly one presentation in five is silent on purpose.
4. If the result shows a real difference across the range, Apply writes a correction. If it does not, the app says so and leaves your EQ alone.

Tones stay quiet throughout, and the EQ is switched off while measuring so the measurement is of you and your headphones rather than of your EQ settings.

Two honest caveats. Without laboratory calibration this measures your ears and your headphones together, so it is specific to that pairing and is not a hearing test in any medical sense — if you are worried about your hearing, see an audiologist. And for most people with ordinary hearing the answer is "nothing to correct", which the app will tell you plainly rather than inventing a curve.

macOS runs a USB DAC at one fixed rate and resamples everything else to it. Unlike iOS it never switches that rate to match what you're playing, so a 5K parked at 96 kHz plays every 44.1 kHz album through a resampler. The row under the volume slider fixes that without a trip to Audio MIDI Setup.

Sample rate row

Pick a rate. 44.1 / 48 / 88.2 / 96 kHz, applied instantly. Match it to what you listen to — most music is 44.1 kHz, most video 48 kHz — and the system stops resampling.

Change what the 5K offers. The device decides which rates it advertises to the Mac, and it can be pinned to a single one; if yours is, macOS has nothing else to offer and the picker will show only that rate. The label at the end of the row says which rates the device is currently offering and changes it. The 5K restarts its USB connection to re-enumerate, so audio drops for a second or two and the app reconnects on its own.

Or let it follow the music. With Auto rate ticked, the app measures what's playing and matches the rate to it: lossless → 44.1 kHz, lossy → the rate you last chose yourself, content that genuinely extends beyond the 44.1 family → 96 kHz. It only acts on a verdict that has held for ten seconds, leaves at least 45 seconds between changes, and never switches while the Soundstage is on (that path resamples anyway). The row narrates every state in plain words — "listening to what's playing...", "lossless — set 44.1 kHz automatically", "lossy — keeping your rate" — so nothing happens silently. Untick it and the rate stays exactly where you put it.

How the detection decides is described under Level; the short version is that it measures the audio rather than asking the player, so it works with Spotify, Apple Music, a local file or a browser tab alike.

The 5K reports its charge over both USB and Bluetooth, and the app surfaces it in three places so you don't have to open anything to know.

In the menu bar. A bolt appears next to the icon while charging. Once the charge drops to 20% the icon is joined by the level itself, so a glance at the corner of the screen is enough. Hovering shows charge, the active preset and which link is carrying the connection.

In the popover. The header pill shows the percentage, orange at 20% and red at 10%, with the state spelled out in the line underneath.

As notifications. Standard macOS notifications at 20% and 10%, and one when charging starts. Each fires once per episode — they won't nag if the reading hovers around a threshold, and they survive a brief Bluetooth dropout without re-announcing the same low battery. macOS asks for notification permission the first time one actually fires, not at launch.

Headphones put the band inside your head. The Stage tab spreads it back out: mid/side width with a brilliance shelf on the sides, an interaural crossfeed (each ear hears a delayed, darkened copy of the other channel — the cue that moves sound out of the skull), a mid-only dialogue lift, and sparse early reflections with a short diffuse tail. Presets for Music, Movie and Theater, plus geometry controls — Distance, Span, Center, Size — and a Night mode that evens out movie dynamics without touching dialogue.

Stage

Crossfeed can be trimmed by band — below 800 Hz, where it anchors the image; between 800 Hz and 4 kHz; and above 4 kHz, where it mostly dulls the treble — and a Balance pair corrects a headphone whose two sides have drifted, in level and in time. Four further stages sit behind the geometry, each off by default and each explained on the pane:

• Loudness — an equal-loudness contour, sized from how far the estimated level at the ear sits below a reference, so quiet listening keeps its bass. It follows a thirty-second average of the listening level rather than this second's chorus.
• Dynamic bass — the 5K applies its curve after the Mac, so a bass boost is a promise the driver has to keep. This stage measures what the device's own curve will do to the low band and eases the loudest passages before they reach it; quiet passages keep the whole boost.
• Impulse response — run a WAV, AIFF or CAF impulse response of your own over the output, with a Mix slider: a headphone correction, a measured room, a reverb. Up to two seconds long, with no added latency at any length; a response the output's buffer size cannot afford is refused with the two ways out named.
• True-peak limiter — the stage ends by holding the output under −1 dBTP, estimating the peaks between samples that a DAC or a lossy encoder redraws. The soft clipper shapes; the limiter guarantees.

The order is fixed: room, crossfeed and balance; convolution; loudness shelves; dynamic bass; soft clipper; true-peak limiter. The Level pane shows what the last three are doing each second.

Unlike everything else in this app, the Stage runs on the Mac, not on the 5K: it processes what the Mac plays on its way to the output device, using a system audio tap (macOS 14.2 or later, and the System Audio Recording permission). The 5K keeps doing its own EQ on-device, so nothing is applied twice. Settings are kept per output device.

When the 5K is on a call over Bluetooth, or a call app is holding the headset microphone, the engine steps out of the audio path and comes back when the call ends, rather than fighting the link while it renegotiates. Separately, a microphone guard watches for anything making the headset's microphone the Mac's default input — which is what drops the link to the 16 kHz voice codec and makes everything sound thin. It can stay off, ask with a banner, or put the default input back on the built-in microphone for you; two reverts of the same device in two minutes and it stops fighting you.

Switching the Stage off takes the processing back out of the audio path immediately. It does not necessarily close the tap, because two other features share it — see The audio tap.

Honesty notes, because this feature category is full of overpromising: it works on the stereo mix — it widens and rooms what is already there. Surround content stays downmixed, nothing tracks your head, and mono content (most YouTube speech) gives Width and Crossfeed nothing to work with — the pane tells you when that is what's playing rather than letting you hunt for a difference that cannot exist.

Live output level, time listened, and how much of it was loud — today and for the previous week, kept 14 days. Useful for the "why are my ears tired" conversation with yourself. Metering rides the Stage engine when it runs, or a listen-only tap (nothing inserted into the audio path) when you switch Track listening levels on by itself.

The meter itself is digital signal level (dBFS). Beside it sits one number about the world rather than the signal: an estimate of the level at the ear, built from a K-weighted measurement of what is playing, the attenuation the 5K itself reports (or the output device's volume when the 5K is not the anchor), and a population figure for the headphone that a per-output calibration slider lets you shift. Two of the three are measured and the third is an assumption, which is why the pane never prints the number without the word "estimate", never prints a decimal, and says what it rests on. Nothing is recorded and nothing leaves the Mac.

The same pane can tell you whether what's playing came from a lossy or a lossless source. It doesn't ask the player — no music app exposes that to other apps — it measures the audio, which is why it works the same for Spotify, Apple Music, a local file or a browser tab.

Lossy encoders discard the top of the spectrum, and where they stop is characteristic: around 16 kHz at low bitrates, around 20 kHz at 320 kbps, while lossless material carries energy to the edge of its sample rate. The verdict appears in plain words with the frequency it measured, and it feeds the automatic rate matching described under Sample rate.

It is evidence, not proof, and the wording says which:

• A lossless file made from a lossy source keeps the original's cutoff and is reported as lossy. That is the correct answer about the audio, even though the file is technically lossless.
• Warm or old masters roll off on their own before any codec would cut them. The app reports "rolls off naturally — can't judge" rather than accusing them, and casts no vote on the rate.
• Quiet or dark passages carry no treble to judge at all, and say so.

Detect stream quality is on by default, and it is one of the three switches that keep a system audio tap open. Untick it and the detector stops; what that leaves running is set out below.

The 5K applies one curve to everything it is sent. Under Presets → Per-app EQ you can give a single app a second curve of its own, applied on the Mac before the audio ever reaches the device — so a podcast player can run a speech-shaped preset while music keeps the curve on the 5K, with no switching by hand.

The section lists whatever is playing right now, plus anything you have already assigned. Pick a preset from your library for a row, or leave it on Default, which means no Mac-side curve at all. Up to eight apps can be assigned at once. Any library preset can be used here, whichever of the device's two EQ banks it was saved for: this chain runs on the Mac and has its own band count, so it is never stretched to fit.

Two things follow from where it runs. It stacks: an app's curve is applied first, and the 5K's own EQ still runs afterwards on the result. And it has to be heard to work, so switching it on with an app assigned inserts the engine into the audio path (the same path the Stage uses) even with the Stage itself off. Switch it off, or leave nothing assigned, and the engine goes back to whatever the other switches asked for.

Four things in this app work on the Mac's own audio rather than on the 5K: the Stage, per-app EQ, the listening-level meter, and stream-quality detection. All of them are fed by one mechanism — a macOS process tap, created global, which captures the output of every process on the machine except this app (excluded so it cannot hear itself). It needs macOS 14.2 or later, and macOS asks once for the System Audio Recording permission.

With apps assigned there is one further tap per assigned app, covering just that app's processes, and the global one then excludes them so nothing is heard twice. A per-app tap that cannot be created is skipped rather than fatal: its app simply keeps playing through the global tap with no curve of its own.

Stream-quality detection is on by default, so the tap is created at launch unless you turn it off. That is also when the permission is asked for — at first launch, rather than the first time you open the Stage. The audio is analysed a block at a time in memory and written nowhere; what survives a block is a verdict and a cutoff frequency.

Four switches decide whether a tap exists at all:

With all four off, the whole engine is torn down, taps included, and the app goes back to being a pure remote control for the 5K.

git clone https://github.com/FrankieMa77/qudelix.git
cd qudelix/QudelixBar
./build-app.sh # current architecture, fast
./build-app.sh --universal # arm64 + x86_64
./make-dmg.sh # universal build, packaged as a DMG

Swift 5.9+ and the Xcode command line tools are all that is required; there are no third-party dependencies.

• Three hosts are contacted, each only when you ask. raw.githubusercontent.com and autoeq.app are reached when you open the Import pane, to fetch the headphone list and the correction you pick. api.github.com is reached only when you press Check in the About panel, to read the latest release number. Nothing is contacted at launch, and there is no background or scheduled check — the update check runs once, when you press it, and downloads nothing.

• The AI preset studio adds a fourth destination, and only if you use it. It contacts exactly one host, the provider you picked in its own menu: one of api.mistral.ai, api.openai.com, api.anthropic.com or openrouter.ai, and never any of the others. It is reached only while you are pressing Generate — choosing a provider, typing a model name, saving a key, picking a preset kind or writing a note sends nothing anywhere. Requests go over HTTPS to that one pinned host, refuse every redirect, and run one at a time with no retry.

• What leaves the machine on a Generate is: the headphone name you typed in the Library field, the preset kind you chose, the note you typed (capped at 200 characters), and — when the AutoEq project has measured that headphone — the stored correction as a list of filter lines this app formats itself from the numbers it parsed. Nothing else. Not your EQ curve, not your presets, not your device, not your output names, and no identifier of any kind. It is your own account at that provider, on your own key, and their privacy policy is the one that then applies to the request.

• The API key lives only in the macOS Keychain, one item per provider under the service name com.qudelixbar.app.ai. It is never written to a state file, a log, an error message, or back into the pane — the studio asks the Keychain whether a key exists to enable its button, and only the request builder ever asks for the bytes, which go into one HTTP header and nowhere else. Forget key removes the item.

• No telemetry, analytics, or crash reporting, and nothing is ever uploaded.

• The Stage and Level features process audio in memory and write none of it anywhere, ever. What they open, when, and how to close it is set out under The audio tap.

• The microphone guard reads which device macOS has as its default input, and in Fix automatically mode sets that default to the built-in microphone — it never opens a microphone, records nothing, and does nothing else.

• Everything the app keeps is a local file, readable only by your user account, and never transmitted. They live in ~/Library/Application Support/QudelixBar/:

  File	What is in it
  stage.json	Soundstage settings per output device, the 14-day listening totals, and the Level, quality and per-app EQ toggles
  profiles.json	Your output-device-to-preset pairings
  last-eq.json	The last EQ curve seen on the device, one per EQ group, what produced it, and which device it came from
  presets.json	The preset library kept on this Mac, the headphone name you typed, and which app is assigned which curve
  ai-research.json	What the AI preset studio has researched, keyed by headphone name — the description it got back, and the measurement it was anchored to. Up to 64 headphones, no key material, nothing about you
  diag.txt	The last 200 lines of an engine heartbeat, for bug reports. It records which phase the studio is in — ai=idle, ai=researching, ai=designing — and never what was asked or answered
  impulses/	The app's own copies of the impulse responses you picked for the Soundstage, named by a scrubbed base name and eight hex digits of the content hash. Copies no output's settings still reference are swept whenever a response is installed or removed

  If one of the five JSON files ever fails to load, it is not overwritten: the app copies it aside as <name>.recovered, carries on with defaults, and leaves the copy for you.

• Outside that folder, ~/Library/Logs/QudelixBar.log holds device packet traces, and rolls over to QudelixBar.log.1 at 2 MB — so there are normally two of it. Since it records raw packet hex it includes the 5K's own Bluetooth address and any preset names stored on it. Separately, macOS keeps the app's preferences in ~/Library/Preferences/com.qudelixbar.app.plist; the app puts two things there — which Bluetooth peripheral it has adopted, as the per-Mac identifier CoreBluetooth issues rather than the device's address, and the AI provider and model name chosen in the studio. Never a key.

• Several of those identify hardware, and are worth a glance before you attach one to a bug report. The packet log and diag.txt carry the names of your audio output devices. stage.json and profiles.json go further: they are keyed by CoreAudio device UIDs, and a UID is generally built from a Bluetooth device's MAC address or a USB DAC's serial number.

• Bluetooth scanning never records the names of nearby devices, only a count of how many were ignored. USB enumeration is narrower but not silent: the 5K's Bluetooth chip vendor is a common one, so the app can meet another device that shares the vendor ID and exposes a vendor-defined HID interface, and the log names that device in the line where it declines to talk to it.

• Only EQ, volume, and preset settings are written to the device — the same things the official app writes. Firmware is never touched.

Sources/qxusb and Sources/qxprobe are diagnostic CLIs, not part of the app. They can destabilise a device if misused — qxusb --noid deliberately reproduces a USB bus-drop failure, and qxprobe sends Bluetooth GAIA commands. Read the source before running them.

• Only the user (headphone) EQ group is exposed, not the speaker group.
• Preset slots show generic names until you name them. Names are limited by bytes rather than characters, so a name in Japanese or emoji runs out of room sooner than its length suggests.
• Crossfeed is shown but cannot be changed here — it lives inside the preset the device stores, and writing it means rewriting that preset.
• A muted band comes back after an app restart as a bypassed band rather than a muted one. The gain is safe on the device and the row reads "Off"; the unmute button is what's missing.
• Profiles match the output device, not your headphones. Two pairs sharing one adapter look like the same output — and some cheap adapters report an identical identity to every other unit of their model.
• Only firmware 3.x is supported; see the table above for what happens on anything else.
• No auto-update mechanism. The About panel will tell you when a newer release exists, but downloading and installing it is manual.

Equalizer

• Band inspector. Click a band's dot to select it; a panel under the graph carries its filter type, gain and a logarithmic Q slider. Double-click a band to clear it. A readout badge in the graph's corner follows the drag, then the pointer, then the selection.
• Flatten and Reset are two buttons. Flatten zeroes the gains and leaves every centre, filter type and Q alone; Reset restores the factory layout for the current mode. Each is one undo step.
• Update writes the curve back into the slot it was loaded from, without a trip through the Save menu.
• The curve and its markers dim when the equalizer is switched off, as the band table already did.
• The gain controls stay on an empty or muted band row.

Presets and corrections

• A preset library on the Mac, beside the twenty slots on the device, holding as many curves as you care to make. A preset is global or bound to one output device, records which EQ bank it was made for, and applies through the same gated, clamped, single-undo path as an import. It lives in presets.json, read and written with the same care as the other state files.
• Headphones field. Name the pair on the end of the 5K. When the AutoEq project has measured that model, a fitted correction is offered once, as a banner and as a line under the field, with the alternative measurements in a menu. A name is looked up once, and only when it changes.
• AI preset studio. Design a preset for the named headphones with an AI provider of your choosing, on your own key: fifteen kinds, from Correction and the published targets to Clarity, Warmth and a V-shape. The draft can be auditioned on the 5K, kept in the library or written into a slot. Silent until you press Generate; the key lives in the Keychain and nowhere else; Correction needs no provider at all when a measurement exists. The Privacy section lists exactly what leaves the machine.
• Per-app EQ. Give a single app a second curve of its own from the library, applied on the Mac before the audio reaches the 5K: a podcast player on a speech-shaped preset while music keeps the curve on the device, with no switching by hand. Up to eight apps, each on any library preset or Default. The section lists what is playing right now, and it inserts the Mac-side engine only while something is assigned.
• Drop a preset file onto the menu bar icon to import it.
• The Presets pane is one scrolling page. Device slots and the library are disclosures that remember whether you left them open, the slot list names the active slot in its header, and the pane stays reachable with the 5K away so the library, assignments and profiles can be tidied offline. Deleting a library preset and overwriting a named slot both ask first. One vocabulary throughout: a slot lives on the device, the library lives on the Mac; Load and Apply bring a curve live, Save to slot... and Save to library... put it away.
• The parser now reads files the way they are actually published: tab-separated, lower case, pass filters without a gain, shelves without a Q, a preamble without its colon. When a file carries more filters than the device has bands, the ones doing the most work are kept rather than the first twenty, and everything a file asked for and did not get is now said rather than dropped in silence.
• Fits are cached on the values the sliders actually take, so re-fitting a shape tried a moment ago costs nothing, and the cache keeps the whole result — the clipping warning and the predicted rating included.

Stage

• Crossfeed by band. Three trims scale the crossfeed amount below 800 Hz, between 800 Hz and 4 kHz, and above 4 kHz. Trims at 100% are a genuine no-op, sample for sample.
• Balance. Level, ±3 dB split between the sides so the loudness holds, and time, up to 0.5 ms, to correct a pair whose two sides have drifted. Applied last, just before the clipper.
• True-peak limiter. The stage now ends in a limiter that holds the output under −1 dBTP by estimating the peaks between samples. Off by default. It runs after the soft clipper, so the last thing to touch the audio is the one stage that can state a ceiling and keep it.
• Loudness compensation. An equal-loudness contour sized from how far the estimated level at the ear sits below a reference, so quiet listening keeps its bass. It follows a thirty-second average of the listening level, and glides to each new setting rather than stepping. Off by default.
• Dynamic bass. The 5K applies its curve after the Mac, so a bass boost is a promise the driver has to keep. This stage measures what the device's own curve will do to the low band and eases the loudest passages before they get there; quiet passages keep the whole boost. Off by default.
• Impulse response. Run a WAV, AIFF or CAF impulse response of your own over the Mac's output, with a Mix slider, per output device. Up to two seconds, with no added latency at any length; a response the output's buffer size cannot afford is refused with the two ways out named.
• The chain order is fixed: room, crossfeed and balance; convolution; loudness shelves; dynamic bass; soft clipper; true-peak limiter.
• Call awareness. When the 5K is on a call over Bluetooth, or a call app is holding the headset microphone, the Mac-side engine steps out of the way and comes back when the call ends. The automatic rate switcher never renegotiates under a call, and the stream verdict measured before the call is dropped rather than left voting.
• Microphone guard. When an app makes the headset's microphone the Mac's default input and the Bluetooth link collapses to the voice codec, the app says so, or puts the default input back on the built-in microphone for you. Off, Ask or Fix. Two reverts of the same device in two minutes and it backs off; a Mac with no built-in microphone degrades to asking.

Level

• Estimated level at the ear. One number about the world rather than the signal, built from the K-weighted loudness of what is playing, the attenuation the 5K itself reports, and a population figure for the headphone that a per-output calibration slider lets you shift. Always marked as an estimate, never printed with a decimal, and the pane says what it rests on.
• The live row shows what the limiter, the loudness shelf and the bass guard are doing this second, and the signal path names each stage only while it is actually in the audio path.

Tune

• Shape. Bass, then presence, then tilt, one axis at a time, three rounds each: twelve comparisons instead of twenty. What you hear is exactly the curve Keep writes. When the pre-gain headroom cannot carry the full ranges with matched levels, the ranges shrink until it can, and the intro says by how much.
• Blind check. Five trials of your curve against a flat one, sides randomised and levels matched. It writes nothing whatever the answer.
• Proper level matching. Pairs are matched on the mean magnitude response from 100 Hz to 8 kHz rather than on the mean of the band gains, so a bass shelf no longer reads as a large change and a treble tilt as a small one. The louder side is trimmed down to the quieter, so matching never makes a session louder than the pre-gain already judged safe.
• A by-ear session earns a verdict before anything acts on it. Too many presses on the silent checks, readings too far apart, or too few of them refuse the result and show the counts the refusal turned on, on the Tones side and the Compare side alike.
• A Compare session survives closing the popover.

The app

• A new shell. The app owns its status item and popover, so clicking the icon no longer leaves a view graph behind each time, and the per-second meter values republish only while the popover is on screen.
• A drawn menu bar icon: headphones in a circle, with EQ bypass, the Soundstage, the stream verdict, battery and a call composed as marks on it.
• Install with a double-click. The disk image carries an installer that copies the app, clears the download flag and launches it, so Gatekeeper asks once, for the installer, and never again for the app.
• The footer shows the version and the source revision it was built from.
• The release build refuses to package a binary that is not universal.

Fixes

• Dynamic bass measured the low band after the loudness shelf had already raised it, so with both on it undid part of the shelf on loud passages. It now measures before the shelf.
• Night mode resumed from an envelope minutes old when switched back on, so it could open several decibels too quiet or too loud. It now starts fresh on every engage.
• The first block after a rate change or an engine start could overrun the realtime deadline at small buffers because eight delay rings were zeroed element by element, twice. They are cleared in one block store, once.
• A library preset with one unreadable band was dropped whole and erased on the next save. One bad band now costs one band. A presets file that exists but cannot be read is left alone instead of being replaced by an empty library, and the AI research cache survives one malformed entry.
• Ending a by-ear session after switching the device between its 10- and 20-band banks wrote the old bank's curve into the new one. The restore is refused across a bank switch.
• Writing an AI draft to a slot could store the previous curve under the new name, because the slot save overtook the coalesced band writes. Pending writes are flushed first.
• Undo now restores the equalizer's on/off switch as well as the curve.
• Bluetooth discovery no longer runs while the 5K is connected over USB.
• The heartbeat file is appended a line at a time instead of being rewritten every fifteen seconds, and it no longer records the name of an impulse response file.
• Nothing is contacted at launch: a saved headphone name is looked up only after the popover has been opened once.
• Dragging a band, hovering the curve, or a device poll that changed nothing no longer redraws the whole popover; the response curve is computed several times faster and the correction search runs on a keystroke rather than on every redraw.
• The audio tap consumed the wrong input buffer when the output device had inputs of its own, so a headset or dock microphone was mixed into what the Soundstage wrote back out and metered as the listening level. The tap's buffer is now taken from the end of the list.
• A tone test could renegotiate the USB rate under itself. Detection sits the session out while the pipeline is muted, and a frozen window is never re-judged as new evidence.
• Hi-res content was downsampled to 44.1 kHz on outputs that stop at 48 kHz. Auto-rate now takes the highest rate at or above 88.2 kHz the output offers, and holds where there is none.
• Outputs that appear a few seconds after login are enumerated again three seconds in, so their saved Soundstage settings no longer look lost.
• Preset and device names are no longer parsed as Markdown in the profile prompt, the microphone banner, or the footer.
• A pane that outgrows its region is clipped at its edge instead of drawing over the footer.
• Imported files go through the hardened reader: no symlink is followed, no FIFO can wedge the load, and a short read is a failed read rather than a truncated document parked as recovered.
• Two bands sitting on the same frequency can no longer be dragged through each other.
• The microphone guard's alerts, the battery alerts and every other notification share one delivery path with fixed identifiers, so a battery dipping in and out of "low" replaces its warning instead of stacking a column of them.

Equalizer

• Drag the curve to shape it. Move a band's dot vertically for gain and horizontally for frequency; hold Option for Q and Shift for fine steps. Hovering shows which band a drag would take. Everything goes through the same gated, clamped path as the sliders.
• Undo and redo for EQ edits, with Cmd-Z and Cmd-Shift-Z and buttons beside Flatten. A drag or an import is one step, not a hundred.
• Auto pre-gain. The editor works out how much attenuation the summed curve needs to avoid clipping in the 5K's own DSP and offers the number. It never moves a value you set yourself.
• Per-band mute for an instant A/B of one band. The gain is kept and restored exactly, and a mute is never committed to the device's flash.
• The requested curve, shown. When an import asks for more than the device can hold, the requested shape is drawn as a dashed ghost behind the live curve with the gap shaded and the worst point labelled.
• Pre-gain is now written to both stored channels. Previously only the first was written, which could leave a left/right level imbalance invisible from the app.

Presets and corrections

• Name preset slots from the app. The name is stored on the 5K, so other software sees it. Saving an imported correction into a slot names the slot after it, unless the curve has since been edited.
• Fit to a target of your choice when importing from AutoEq, with bass and tilt adjustment on the same fit. Targets are grouped by form factor. The fit is asked to stay within the device's limits, so nothing is clamped on arrival; the published preset is still offered as-is.
• Predicted preference rating for over-ear and on-ear corrections, from the published AES model, scored on the error that survives the bands this device can actually run. In-ear gets no score: those coefficients could not be verified from a primary source.
• Import from the clipboard. Paste a filter list as published sites print it.
• Bandwidth-aware fitting. When stream-quality detection has measured where the source stops, an import can be told to stop correcting there.
• Profiles. Pair an output device with a preset and be offered the switch when that output becomes active. It always asks the first time; mark an output automatic once you trust it.
• The EQ snapshot used to repair a device is now kept per EQ group, tagged with the device it came from, and never restored onto a different device. A damaged snapshot is parked rather than overwritten.

The device itself

• Device settings drawer: channel trim, a volume ceiling, the DAC reconstruction filter, power and battery-care state. Crossfeed is shown but not settable.
• The parser now reads the codec, output jack, gain, mute, charger and low-battery fields it used to skip. The low-battery alert takes the device's own flag as a trigger, and the header can say "plugged in, not charging" when battery care is on rather than implying a fault.

Everything else

• Signal path inspector in the Level pane: five rows from source to output saying what alters the audio and what passes it through.
• About panel with the version and the exact source revision, and an update check that runs only when you press it.
• Accessibility names for every icon-only control.
• Copy and Reveal buttons on the diagnostics log.
• The README and privacy section now list every file the app writes and state that a system-wide audio tap is created at launch by default, with the three switches that govern it.

Fixes

• Bluetooth stopped retrying a switched-off device every 18 seconds for the whole session. It backs off to once a minute and logs it once.
• The by-ear tuner read headroom off the largest single band rather than the summed response, so it under-read on exactly the curves with the least headroom. Tune results derived on top of an import are worth re-running.
• The by-ear macros and the tone-test correction were applied by band index, so in 20-band mode they landed on the wrong frequencies. Both now map by frequency.
• A/B trials were being written to flash and pushed into undo one band at a time. A trial now reaches neither.
• Quitting with the 5K unplugged could save a flat curve as the last EQ and write it over the real one on the next connect. Snapshots now require a device that has been read.
• Device reports stopped arriving during any slider drag and landed in a burst afterwards, defeating the echo windows. The HID source now runs in common run-loop modes.
• Recording of the listening history now follows its switch alone. It used to accumulate whenever the engine ran for stream detection.
• Auto-rate refuses to renegotiate the 5K from audio that was going to a different output device.
• A tone test that measured nothing at a frequency says so instead of reporting typical hearing and writing a "+0 dB" correction through the gap.
• The Compare session's reliability check no longer flags a listener who correctly answered "same" on identical pairs.
• The Stage soft clipper is anti-aliased.
• Hardened runtime on the bundle. The audio-recording permission is asked once more on first launch of this version because the grant is keyed to the signature.
• Two audit rounds of crash, resource and input-validation fixes: NaN and non-finite sample rates, unbounded loops on hostile buffer descriptions, throttled logs, bounded caches, state files that refuse anything but a regular file.

• Stage: a soundstage for headphones, applied on the Mac.
• Level: live output level and a 14-day listening history.
• Stream quality detection, with optional automatic sample-rate matching.
• USB sample rate control, including the rates the device offers the host.
• Battery indication in the menu bar, with low and charging alerts.
• Switch between 10-band and 20-band EQ from the app.
• EQ settings persist across device restarts.

• Control over Bluetooth LE as well as USB.
• Tune tab: blind A/B comparison and tone thresholds, by ear.
• Fixed the EQ mode being misread on connect.

• Hardened device, log and network handling; release checksums published.

• First release: USB control, live status, volume, 10-band parametric EQ, 20 preset slots, AutoEq import, diagnostics.

Bug reports, ideas, and pull requests are all welcome — open an issue or a PR.

Especially useful right now:

• Firmware other than 3.1.8, or any 5K that behaves oddly — the diagnostics panel and ~/Library/Logs/QudelixBar.log capture everything needed.
• Intel Macs. The binary is universal but has only been run on Apple Silicon.

MIT. Use it, build it, fork it; no warranty is given.

• devicePEQ — prior open-source work on the Qudelix HID protocol
• AutoEq by Jaakko Pasanen — the headphone correction database