StemDesk Release Notes
The current version is v1.0.53, on Windows, macOS and Linux. Updates are never installed automatically: the app's Settings → Check for Updates shows you what changed and offers the download.
v1.0.53
You can now hear Auto-Mix on your own stems in the app on every plan. The 17 plug-ins are rebuilt at 1.0.53.
Purchases made before 2 November 2026 keep the refund terms in force when you bought (14-day refund for any reason).
Until our checkout moves to Stripe Managed Payments, purchases are processed by Fulks, Inc. through Stripe, and the references to Link as merchant of record in the Terms and EULA apply from that move.
New
- Hear Auto-Mix on every plan. Free and Pro render an Auto-Mix to a private preview you can play and A/B in the app. Saving or exporting the Auto-Mix files is still Studio ($179, one-time).
- Clearer upgrade path. The upgrade window shows what each plan includes and the prices (Pro $99, Studio $179, one-time). The licence screen has a "Buy a license" link, and a quiet "Free" chip in the title bar shows your plan.
- Update note at startup. StemDesk checks once at startup and shows a short "Update available" note you can dismiss. Turn it off in Settings. Updates are never installed automatically.
- Genre is asked, not guessed. A new profile asks for the genre and remembers your last choice. Settings can be set to "None - ask me".
- Make a song is now a Labs option, off by default (Settings → Labs). It needs the ACE-Step download (about 8 GB).
- Honest labels. A result that came from built-in rules rather than the AI model carries a "Heuristic (no AI)" badge. The Figures-used list under a critique shows which measured numbers it relied on.
- Stems that did not go through their plug-ins are named. After an Auto-Mix, StemDesk lists every stem that was not rendered through its plug-ins (rendered with a simplified built-in chain, only partly rendered, or not processed at all), says why and what to try. The list is saved with the session and shown again when you reopen it. A stage your plan does not include is not reported as a failure.
- Silent stems are left alone. A stem that is completely silent passes through at its planned level with no processing and no failure message, and the result says which stems were silent.
- The left and right files of a stereo split are mixed as one track. Tracks named like Vox.L and Vox.R (a stereo file split by a DAW) are recognised as one stereo stem. They are classified together, get one fader row, and are mixed like the same audio as a single stereo file; before, one side of a split lead vocal could come out far quieter than the other. If you choose different roles for the two halves, StemDesk applies neither and asks you to choose again.
- Lyrics detection leaves out what it cannot read. Passages the speech detector cannot read reliably are left out of the lyrics and StemDesk says how many. The same stem gives the same words every time you run it.
- The mix leaves your Presets folder alone. A mix does not rewrite the preset files in your session's Presets folder. What the mix applied is kept in a separate Rendered folder, and the exports use what the mix actually applied.
- Install presets installs what the Auto-Mix used. The presets you install carry the values StemDesk derived at the intensity you last heard, for the stem before any gain change (a note in each StemDesk preset says so). Stems whose names end alike, such as bass and sub bass, each get their own preset. If the presets cannot be derived, the app tells you to generate them again.
- Better progress. Auto-Mix shows "Step n of N" with the time elapsed. Stem splitting shows the overall percent and time left. A second split of the same file with the same settings reuses the stems from last time and says so.
- Stem separation settings. Settings → Stem separation: quality Standard or Extra (slightly cleaner, about twice as slow), how much separated audio to keep (off, 1, 2, 5 or 10 GB), and a Clear stem cache button.
- Tune lead vocal (off by default). An Auto-Mix option that pitch-corrects the lead vocal to the detected key. Check the key first. It works with the StemDesk plug-in set only, and the status line tells you the key it used, or that it was skipped because no readable key was found.
- Multiband on a single stem (optional). In the Chain Editor, Add multiband puts a gentle four-band compressor on one stem. Its thresholds are set from that stem's own level. StemDesk plug-in render only.
- Advanced analysis. Settings → Advanced analysis. A gentle role-based EQ is used when the analysis cannot measure a track, and EQ plus compression is used when the plug-ins cannot load. Both are on by default. The experimental options are off by default.
- The model download names its licence. The dialog that offers the language-model download now says it is Google Gemma 4, released under the Apache 2.0 licence, with a link. The licence list in Settings matches.
- Plug-in settings travel with REAPER and Ableton sessions. With the StemDesk plug-in set, a REAPER or Ableton (.als) export now opens with each track's plug-ins already set to this session's settings (checked in REAPER and in Ableton Live 12). DAWproject files include the plug-in settings and validate against the DAWproject format; they have not yet been tested in Bitwig or Studio One.
- Automation in Ableton. Ableton (.als) exports now carry the volume-ride, duck and section automation as a track-volume lane, like the REAPER export already did (checked in Ableton Live 12).
- The DAW export now follows the render stage for stage (StemDesk plug-in set). The plug-ins on each track are the ones the Auto-Mix played, in the order the Auto-Mix played them, including the order you set in the Chain Editor (before, the export used a fixed order). The Stereo Imager, the last stage of the chain, now comes along as two plug-ins in series (the render runs it twice), and the vocal Cleanup stage comes along on the vocals it ran on. A stage the render skipped or could not run is not in the export.
- Mono tracks are exported as the stereo the mix hears. The render widens a mono track to stereo before its effects, so the export now saves that stereo version next to the project (in a folder named after the project) and the track plays from it; your own files are not touched. Volume rides are applied after the widening in the DAW and before it in the mix; the export note gives the measured difference, in dB, for each track that has a ride. Stereo tracks are exported as before.
- A DAWproject export carries the planned settings when the exact settings cannot be derived again. When the settings StemDesk mixed a track with cannot be re-derived (an older session, a session without its analysis file), that track is exported with the planned settings at its own fader level, and the export note names the tracks. Run the mix again, then export, to get the exact settings. Nothing is exported silently different: the note is shown every time this applies. A track whose stamped settings cannot be saved by the mix, or whose Stereo Imager settings are not in the Library, is exported the same way, so its gain is never counted twice; the note names it and says to run the mix again.
- Library and Export folders are only ever replaced by StemDesk's own files. Anything you put in them stays: a file of yours whose name matches a new StemDesk file is kept as `name (yours)`, never deleted with the old folder. If StemDesk cannot read its own list of what it wrote there, it moves every file in that folder to a `.quarantine` folder in your session (nothing is deleted) and the mix result says where.
- Multiband and the lead-vocal tuner come along. A stem that had them on in the last render gets them in the exported chain, in the same place.
- Each track's gain is placed where the mix applied it. In REAPER the gain a track got in the mix is on its clip, before the plug-ins, and the plug-ins hold the exact settings applied at the intensity you rendered; leave the fader where StemDesk set it. DAWproject has no clip gain, so there the gain is on the fader, after the plug-ins, and compressors and saturators may react slightly differently. A track that needs more than the fader's +6 dB is limited to +6 dB and named in the export note.
- The export tells you what it could not carry. If a mono track cannot be converted to the stereo the mix hears, it is exported as plain mono and will sound narrower than the mix, and the note names it. If a track's Stereo Imager settings file is missing or unreadable, the imager is left out of that track and the note says so.
- The export stops when your stems or roles changed after the analysis. If you replaced or edited a stem, or changed a role, since you analyzed, the export (and the Pro Tools build) stops, names the stems, writes nothing, and asks you to analyze again or cancel,. If the session's stems folder cannot be found at all, the message names the folders that were checked. The same goes for presets made from a different analysis.
- Updated terms. You will be asked to accept updated terms once (Terms of Service, EULA and Privacy Policy, legal documents version 2.10.0: seller details, tax, EU withdrawal rights and the startup update check).
- Smaller improvements. Tempo ×2 and ÷2 buttons; an Open session button on every empty screen; the model list is ordered for your graphics card; one-letter processor chips have tooltips; the transport keeps the song length after Stop; the debug log is capped at 10 MB; the real-time meter stops when the window is hidden; autosave only writes when something changed; your plug-in folders are scanned once and remembered between launches.
Mastering
- Hard-hitting genres are mastered up to 1.5 LU under the platform target on purpose, to keep drums punchy. A very peaky mix can land further under when the true-peak limit leaves no room, and StemDesk shows by how much.
Plug-ins
- Tooltips in the Channel Strip, Reverb and De-Esser are translated, and De-Esser DETECT is one undo step.
- New EQ bands start at a flat Q (0.71): a High Pass or Low Pass no longer humps by 1.2 dB at its corner. Saved sessions and presets keep their own Q.
- New Multiband Compressor instances keep the top octave flat (the built-in second oversampler is gone) and report up to 23 samples less latency. A session saved by an earlier version with the sound update may sound very slightly different, and a one-time notice says by how much; sessions saved before the sound update are unchanged.
- Transient Shaper: both channels now share one detector, so a hit that is louder or only on one side no longer pulls the stereo image. This applies to new instances and to sessions saved with the sound update; older sessions are unchanged.
- Reverb: the High-pass and Low-pass filters now follow automation smoothly and on time, instead of trailing it by a fraction of a second.
- Reverb: Decay Hi now does what its label says. At x0.25 and x4 the high-frequency tail is clearly shorter or longer than before, in every mode. Saved sessions with Decay Hi away from x1 show a one-time notice; sessions at x1 sound exactly as before.
- Multiband Comp: each band now uses the same improved compressor as the Compressor plug-in (cleaner Punch, Knee in dB, Drums Attack and Release now mean what the knob says), and loud settings no longer clip flat at +12 dBFS.
- The high-pass filters in the compressor, gate, de-esser, EQ, stereo imager, multiband compressor, fattener, delay and reverb now run in double precision, which removes low-frequency filter noise at high sample rates. A saved session whose settings would change by more than 0.01 dB shows a one-time notice naming the plug-in and the amount; every other session sounds unchanged.
- Saturator: Tape now has extra anti-aliasing at 2x, 4x and 8x, and at 1x from 88.2 kHz up. A saved session that would sound different shows the one-time notice.
- The automatic de-esser now works on lead vocals (StemDesk plug-ins). It sets each lead vocal's reduction from the vocal itself. On lead vocals it had never been tuned on, it met all four of its targets on 92 % and 94 % of the vocals where that is possible (it stays off on the others): it took 2 to 4 dB off the S sounds, without pumping, and without dulling the rest of the vocal by more than 0.5 dB, in two fresh test sets measured on the plug-in's real output. A session analysed before 1.0.53 keeps its old de-esser setting until you analyse it again.
- The Detect button on the de-esser now reads the whole vocal it has heard, and says so when it has not heard enough; on a fresh set of vocals it landed the right setting on 82 % of tries.
- The Stereo Imager sounds slightly different near its crossover points. When Width is not 100%, the sound around the crossover frequencies can differ by up to about 0.2 dB from earlier versions, so a mix you render again can change by that small amount.
Fixed
- Security fixes. This update fixes a local security issue in StemDesk's plug-in helper and installer. 1.0.53 also checks that every update notice really comes from us (a digital signature). Versions 1.0.52 and older cannot check this, so please update.
- At 250% display scaling in REAPER on Windows, a plug-in editor now fills its window instead of drawing small in a corner.
- Plug-in tooltips with dashes and accents no longer show garbled characters, and they appear in your language.
- The first Analyze of a session can no longer leave the channel strip waiting for ever: it now answers within two minutes, with an error if the analysis hung.
- The Mac AI runtime folder spelling is consistent. An existing folder is moved once, safely.
- Deleting a paid plug-in bundle through the helper now leaves a line in the helper's own log.
- A REAPER, Ableton or DAWproject export now removes a mono stem's DC offset exactly as the Auto-Mix does, and the export notes name those stems. So if you re-export an older session, a stem with DC offset is written to a different file than before, and that file is now the correct one.
- A crash, a power cut or a full disk while StemDesk saves a mix report, a preset, a curve export, a diagnostic report, a lyrics file or an export file can no longer leave a half-written file under the final name; the previous file stays.
- If the licence file cannot be read for a moment (an antivirus or sync tool holding it), the Auto-Mix does not assume the Free plan when it checks for missing plug-ins; it names the missing plug-in.
- A song last mixed with a build from before 1.0.53, whose compressor setting cannot be carried over exactly, gets its compressor left off, with a note that says to analyze the track again.
- If a session's saved settings file cannot be read, the session still opens, says so, and keeps your analysis. If you switch to another song while presets are being generated, StemDesk says the presets were saved with the first song only when they were.
- If StemDesk cannot analyze or join the left and right files of a stereo split, the analysis says so, and each side is classified on its own.
- The per-stem mix files Auto-Mix saves now line up with your original stems; before, they could be about 5 ms (240 samples) late.
- Auto-Mix never guesses the Intensity: a render that has none stops and asks you to set it.
- Stems with long names that start the same (for example the left and right files of a split) no longer pick each other's presets; before, both could be rendered unprocessed. If two presets still match one stem, StemDesk says so.
- Open preset folder now says why when the folder cannot be opened.
v1.0.52
Please update. 1.0.50 and 1.0.51 could open with an unresponsive main window. 1.0.52 fixes it. Please download it from the website. The 17 plug-ins are rebuilt with the fixes below.
Sound
- Drum Tuner: no more double hits. Each drum hit now plays once, in the Drum Tuner and in the Channel Strip's tuner.
- Cleaner bus compression. The glue compressors on the master and group buses no longer react to the low end, so kick and bass stop pumping the rest of the mix.
- Fattener level. With depth at zero, the Fattener no longer doubles the dry signal, so the level stays where you set it.
- Compressor timing in Pro Tools. When you change the Compressor's lookahead, the timing it reports to your DAW now matches what it really does, so tracks stay lined up.
- Brighter vocals tamed on sparse songs. On songs with one lead vocal over a sparse beat, the master no longer adds as much 3.5 kHz presence.
- The right level when stems are skipped. If Auto-Mix has to skip a missing or unreadable stem, it now sets the mix level for the stems it actually uses, so the mix no longer comes out too quiet.
Safety
- Verified updates. The app now downloads the new installer itself and checks its fingerprint against our release list before it shows it to you. If it does not match, it is deleted. If it cannot be checked, the app sends you to the download page on the website.
- Recycle Bin. Deleting a session now moves it to the Recycle Bin, so you can get it back. If that is not possible, the app tells you it was deleted for good.
- Damaged files are kept. If a session's saved state is damaged, StemDesk sets it aside instead of writing over it.
- Errors are reported. If an analysis file is damaged, Auto-Mix stops and says so instead of carrying on without it.
Fixed
- Chinese and Korean text now displays correctly everywhere, including the new messages.
- A few messages that only showed in English are now translated into all 25 languages.
- If a session has no genre, the app now tells you it used the neutral plug-in chain.
- Tiny EQ corrections that are too small to apply are now listed instead of silently skipped.
- A tempo outside 20–300 BPM in a session file is replaced with 120 and noted, instead of breaking tempo-synced effects.
- Saving large audio files no longer makes the app pause.
- Imported older mix profiles keep settings for tracks that are not in the current session, and the app tells you how many.
- The Plug-in Manager is steadier on shared PCs and never removes the free plug-ins.
- The Channel Strip handles very large host buffers without glitches.
- The Compressor uses less CPU. It sounds exactly the same.
- On Linux, a downloaded update is ready to run.
- On Linux, Analyze is fast again. On some computers it could take twenty minutes or more.
- On Mac and Linux, the preset files for the StemDesk plug-in chain are now created, and saving a session before its first run no longer shows an Invalid path error.
- Vocal AI now finds the words in a vocal on every computer; it was asking for a speech model that was not included. If it cannot read the words, the app now says so on screen.
- On a fresh Mac install, exporting FLAC, MP3 and AAC works again. Installs that missed it repair themselves the next time StemDesk starts.
- In a Pro Tools session, tracks are only put in your session's order when their names really match.
- A rare Could not save mix targets error after an analysis is fixed.
- The one-time download notice for stem separation no longer shows on every 6-stem run.
- The website now lists exactly what StemDesk leaves behind when you remove it on Mac and Linux.
v1.0.51
A safety and polish update on all three platforms, plus one new option in Auto-Mix. The 17 plug-ins are rebuilt with the fixes below.
New
- Group Bus (opt-in). A new Auto-Mix checkbox that glues related stems together at bounce time. It is off by default, and with it off your mix is exactly the same as before.
Fixed
- Shared PCs. Signing in as a second person on the same PC no longer removes plug-ins the first person still owns. If plug-ins are kept for that reason, the app says so instead of saying they were removed.
- Settings recovery. If your settings file is damaged, StemDesk now restores what it can instead of resetting every setting to the defaults.
- No more freezes on big session clean-up. Comparing large session folders now runs in the background, so the app stays responsive while it works.
- Honest uninstall. If any plug-in or data folders are left behind, the uninstaller now lists them instead of reporting success.
- The right AI model on 8 GB machines. 8 GB graphics cards and 8 GB Apple Silicon Macs are now offered a model that fits, instead of one that could run out of memory.
- Mastering target note. If a custom loudness or true-peak target has to be limited, the app now shows a note and puts the value it used in the field. Before, it changed it silently.
- If your sessions folder is not usable, the app tells you and explains how to choose a subfolder.
- If the licence revocation list cannot be read, StemDesk uses the last good copy it saved, so revoked keys stay blocked.
- On Mac, installing the Pro Tools plug-ins checks that every file is there before it copies anything, so you never get a partial install. The notice about a folder that cannot be written to now appears again if the problem returns.
Plug-ins
- Saturator now reports its timing to the host correctly when you change oversampling.
- Multiband Compressor stays correct at very high sample rates.
- A rare click when a convolution kernel changes quickly is gone.
- Equalizer presets that leave bands unset now behave the same everywhere.
v1.0.50
Pro Tools on the Mac is here, and a full audit of the app and plug-ins fixed the rest. The 17 plug-ins are rebuilt with the sound fixes below.
New
- Pro Tools (AAX) on macOS. All 17 StemDesk plug-ins are now available as AAX in Pro Tools on the Mac, signed by PACE and Apple-notarized, with no iLok needed. They install next to the VST3 and Audio Unit versions; if you install Pro Tools after StemDesk, restart StemDesk to add them.
- Track clashes card. Shows the track pairs that mask each other, with a suggested cut.
- Volume and input-trim faders can now be moved with the keyboard.
- The Tuner and Vocal Cleanup can be switched off per Channel Strip. When off they are not processed and add no latency: a new strip reports about 3 ms instead of about 64 ms. Saved sessions keep both on.
Improved
- Width suggestions fit the genre. EDM mixes are flagged below about 0.30 width; trap and drill below 0.15.
- Clearer analysis warnings. A stem with bad samples is still cleaned but is now counted and shown. A track that cannot be identified is named and mixed as a neutral effects track, and one failing track no longer stops the rest of the analysis.
- Printed-effect detection no longer mistakes names like "template" or "Delayed_Vox" for reverb or delay returns.
- The level card always shows the target against the loudness reached, and says why when the master lands more than 0.5 LU away.
- Mix Coach requests run one at a time, so they no longer collide.
Plug-ins
- Synced Delay follows the project tempo.
- New instances of Saturator, Fattener, Reverb, Compressor, Channel Strip and De-Esser have the sound fixes. Saved sessions keep their existing sound.
- Fattener mode changes fade instead of clicking, and Drum Tuner timing is tighter; existing sessions using these may sound slightly different.
- De-Esser in Split mode with an external sidechain now works on the program's own band. Sessions that use it will sound different.
- Undo, redo and loading a state no longer cause a silent gap on a live track.
- Recalling a 1.0.49 preset while playback is running can cause the same brief gap as moving that knob: Saturator about 6 ms, Multiband Compressor up to about 80 ms, Channel Strip up to about 60 ms.
Fixed
- A damaged licence file is backed up and restored, and a torn file cannot strip paid plug-ins. Deactivating while a DAW is open finishes the plug-in removal at the next start.
- Renaming a session updates the file paths inside it. Saving a preset no longer replaces a different preset: the new one gets a number.
- If your saved sessions folder is missing, the app now tells you.
v1.0.49
Maintenance update on all three platforms. The plug-ins stay at 1.0.46: nothing about how StemDesk analyses, mixes or sounds has changed.
Fixed
- The AI engine never flashes a console window. On Windows, the local AI engine's background process now starts fully hidden — no black terminal window appears when the app launches.
- The transport logo stays put during playback. The STEMDESK logo plate in the transport bar no longer shifts as the loudness readout's numbers change.
v1.0.48
Maintenance update on all three platforms. The plug-ins stay at 1.0.46: nothing about how StemDesk analyses, mixes or sounds has changed.
Fixed
- Windows: no more console windows. Analysis and Auto-Mix no longer flash black terminal windows on screen while they run.
- Analyses auto-repair a corrupted cache. If an analysis was interrupted at exactly the wrong moment, every later analysis could crash silently at the first stem. The app now detects that crash, clears the damaged cache and retries the step once.
- Auto-Mix steps are watchdog-based. A step is stopped only when it stops producing output, instead of at a fixed per-minute limit — a genuinely slow machine no longer loses Auto-Mix to a timeout unrelated to its progress.
v1.0.47
Maintenance update on all three platforms. The plug-ins stay at 1.0.46: nothing about how StemDesk analyses, mixes or sounds has changed.
Fixed
- The plug-in self-check suite now passes on Windows. The checks covering the Vocal Tuner's tuning engine ran out of thread stack on that platform after the 1.0.46 audit widened the tuning buffer; they now allocate the way a host does. Shipped code was never affected — a host always puts a plug-in on the heap, not the stack.
v1.0.46
The Auto-Mix now sounds the same with the StemDesk plug-ins as with FabFilter and lands on the loudness you set, and a full audit of the app and plug-ins fixed the rest. Plug-ins update to 1.0.46.
Improved
- Delays sit on the beat. Auto-Mix delays use musical note values (dotted 1/8 or 1/4) at the song's real tempo, and are left off a vocal that already has delays printed in.
- StemDesk plug-in mixes match the FabFilter mix in tone, and both land on the same master loudness.
- Drums knock harder, and the default strength is 30.
- Long songs keep every plug-in. Songs over about 12 minutes used to fall back to a basic compressor.
- The A/B reference player now matches loudness while you compare, and loading or clearing a reference works cleanly.
Plug-ins
- Bypass stays in time with your DAW's delay compensation (Vocal Cleanup, Vocal Tuner, Drum Tuner, De-Noise, Saturator).
- Saturator at 2x/4x/8x: the dry signal no longer carries Input Gain. Saved sessions with Input Gain set and Mix below 100% will sound slightly different.
- Transient Shaper no longer clips its dry signal, plus stability fixes across the set.
Fixed
- Cancel works every time, and quitting closes everything the app started.
- Switching sessions quickly no longer mixes up results between them.
- Presets and exports are written safely, so an interrupted run never leaves a broken file.
- Many smaller fixes from a full code audit.
v1.0.45
Everything you set in the app now lands the same way in the Auto-Mix file, the in-app player and the Channel Strips in your DAW.
Improved
- Delay, drum tuning and per-track limiting are now in the Auto-Mix file, rendered by the Channel Strip with the same settings your DAW gets. Turning off a track's reverb dot also takes it out of the shared reverb.
- Intensity, input trim, the master fader and the chain order now reach the Auto-Mix file and your DAW, not only the in-app player.
- Changing a track's role re-plans its gain, pan and processing.
- The track list fits laptop screens: every column, from gain to the processor dots, is visible at 1280 × 720 and up.
Fixed
- The drum tuner no longer transposes bass or 808 lines. It now tunes only kicks and snares, and only when it reads the drum's real pitch.
- Nudging a track's gain no longer throws away its analysed level in the Auto-Mix file or in exported DAW sessions.
- The pan knob on stereo tracks now reaches the Auto-Mix file.
- The in-app player plays each track at its analysed level, and unticked tracks are silent.
- A session keeps its key when reopened, and its last Auto-Mix result can be opened again.
- The split "Instruments" stem no longer asks you to confirm its role.
v1.0.44
Fixed
- AI suggestions can be accepted again. The Accept All / Reject All bar now appears right above the track list, and suggested tracks stay highlighted. Reject All puts the track's value back on screen.
- The chorus lift shows and plays in the app right after Auto-Mix, without reopening the session.
- The level card states how much quieter the master really is when a hard-hitting genre is kept under the platform level on purpose.
- The Auto-Mix results panel no longer covers the status bar.
v1.0.43
Vocal AI now really rides the vocal, and what you hear in the app and in your DAW matches the Auto-Mix file.
Improved
- Vocal AI rides the lead vocal word by word. The speech model it needs now ships with StemDesk; before, it was never installed, so Vocal AI quietly used a basic rider. Soft words come up and loud words come down, toward the vocal's normal level.
Fixed
- FabFilter masters reach the loudness target. Pro-L 2 could land slightly short; it now checks and corrects.
- The level card explains a quieter master: hard-hitting genres are kept slightly under the platform level on purpose, so the drums keep their punch.
- The chorus lift plays everywhere: section level changes were only in the Auto-Mix file; the in-app player and the plug-ins in your DAW now play them too.
- Bypassing a process on a track now affects the Auto-Mix file, not just the preview and your DAW.
- The AI report's suggested level change now appears in the track table, and no console windows flash while StemDesk works in the background.
v1.0.42
A reliability release, after a full bug hunt and testing against many real-world folder layouts.
Fixed
- Auto-Mix applies compression and saturation again, and Split a 2-Track works again. In 1.0.41 the compiled engine could not start those steps, so they were silently skipped.
- Stems folders: standard 4-stem folders keep their drums; an MP3 reference, macOS ._ files or an empty file no longer stop a session; your own multitrack folders keep every track; each 2-track split gets its own folder.
- Plug-in installs: a second Windows user or a missing license file no longer removes the paid plug-ins, and one stray file no longer blocks installing them.
- Plug-in connection: reconnecting keeps your settings, rides for long songs arrive reliably, and AI prompts are not lost on a brief disconnect.
- Many smaller fixes: file paths containing #, ? or %, opening StemDesk when Windows cannot find your Documents folder, Batch Album songs in subfolders, license keys pasted with typographic dashes, and a faster start-up.
v1.0.41
Protecting what makes StemDesk mix the way it does.
Changed
- The mixing engine now ships as compiled code instead of readable scripts: gains, pans, rides, genre profiles and preset generation. Same results, measured slightly faster. Nothing about how you use StemDesk changes.
- The license check ships as bytecode and is verified at more than one point. Still fully offline; there is no activation server.
- Developer tools are no longer reachable in the installed app. The plug-ins are unchanged from 1.0.40.
v1.0.40
A clearer, more honest look across all 17 plug-ins.
Added
- See what Delay, Reverb and De-Noise are doing. Delay shows its echo pattern: the dry hit and each repeat, alternating left and right with ping-pong. Reverb shows its pre-delay and how the tail fades away. De-Noise shows a scrolling view with the noise it removes in amber.
- Redo now sits beside Undo at the top of every plug-in.
- How loud is my file, and which one do I use? After Auto-Mix, a card shows how loud the Master came out and the peak of the Mix, and says which file is for uploading and which is for a mastering engineer. The mastering menu is now called Final loudness, and the manual explains what "−6 dB headroom" advice really means.
- Big sessions: above 40 tracks, the app suggests grouping tracks in your DAW first.
Fixed
- Panning in big sessions. Tracks named "BG Vocal", "Backing Vocal", "Harmony", "Double" or "Ad-lib" were treated as the lead vocal and all sat in the centre; they are now spread left and right. Several tracks of the same instrument, such as synths or keys, no longer pile up on one side.
- VU meters read correctly. The analog dials showed peak level, so they sat pinned on any normal track. They now read average level like a real VU meter (0 VU = −18 dBFS), and the needle lines up with the printed scale.
- Long decimals in readouts. The Vocal Tuner's correction, the Delay time at some tempos and Channel Strip frequencies now show rounded values.
- Easier to read. Larger, brighter knob names, highlighted values no longer smaller than the rest, each knob's full range visible, the same "StemDesk" title on every plug-in, and taller windows for Vocal Tuner, Drum Tuner and De-Noise.
- The Vocal Tuner's AI screen now says AI isn't available there yet, instead of staying blank.
v1.0.39
Chinese, Japanese and Korean text in the plug-ins on macOS.
Fixed
- Blank buttons in Chinese and Japanese on macOS. Inside a DAW such as GarageBand or REAPER, the plug-ins' Save, Undo and Redo buttons showed no text in Chinese and Japanese (Japanese showed only some of its characters). The plug-ins now use the Mac's own Chinese, Japanese and Korean fonts, so every character appears. Windows and Linux were not affected.
v1.0.38
The website, checkout and update notice now speak your language too, and no plug-in word is cut short in any language.
Added
- stemdesk.ai in all 25 languages: the home page, the user manual, these release notes and the source-code page. The site opens in your browser's language; the language menu at the top changes it.
- Checkout in your language: the Buy buttons open Stripe's checkout in your language wherever Stripe offers it (20 of our 25 languages; Arabic, Hebrew, Hindi and Ukrainian buyers get Stripe's choice for their browser).
- The update notice in your language: from this version on, Settings → Check for Updates describes the new version in the app's language.
Fixed
- No more cut-off words in the plug-ins. In some languages the Save, Undo, Redo and Reset buttons and the Channel Strip's Analyze status (for example CONFIGURED or NO APP REPLY) were too long for their buttons and showed only their first letters. These buttons now size themselves to their text. The De-Esser's REM button was cut even in English; it fits now.
- Every build now opens each of the 17 plug-ins in all 25 languages and fails if any button or label would cut its text.
v1.0.37
StemDesk in your language: the app, all 17 plug-ins, the Windows installer and the in-app help now speak 25 languages.
Added
- 25 interface languages: English, Arabic, Chinese (Simplified and Traditional), Danish, Dutch, Finnish, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese (Brazil), Russian, Spanish, Swedish, Thai, Turkish, Ukrainian and Vietnamese. StemDesk follows your system language; change it in Settings → Language.
- Right-to-left layout for Arabic and Hebrew. Meters, faders, pan and the waveform keep their usual left-to-right direction, as in every DAW.
- The AI mix critique and Mix Coach answer in your language.
- The plug-ins show their menus, tooltips, dialogs and status in your language. Knob names, mode names and parameter names stay in English — they are the standard audio terms, and parameter names are what your DAW stores in automation and sessions, so sessions stay compatible across languages.
- Bundled fonts for Arabic, Hebrew, Hindi, Thai, Cyrillic and Vietnamese, so every language displays correctly on every system.
Fixed
- The DAW Compatibility help page shows its table again. Opened from disk, browsers blocked it from loading its data, so it showed "Failed to load". It now also lists the Audio Unit format on macOS.
- A ride-curve update can no longer free memory on the audio thread in the Channel Strip (the last two paths that could are closed), and the audio thread no longer takes a lock to read the ride curve.
- The loudness meter's Reset button no longer sits off-screen on laptop-width windows.
- The first-session tip names the button that is really there (Start Mix Session).
Also
- Every plug-in is now tested on each build for heap allocation inside the audio callback, and the plug-in ↔ app connection is tested end to end (authentication, large messages, a stuck plug-in, app restart). The macOS build now runs the same packaged-app render check as Windows and Linux.
- The complete corresponding source code for the FFmpeg libraries shipped on Windows and Linux is available at stemdesk.ai/source, and the licence file lists every component built into them.
- Legal documents (EULA, Terms, Privacy Policy) stay in English; the app says so on their screen.
v1.0.36
Attribution release: the third-party licence file that ships inside the app now includes the AI components the app downloads after install.
Added
- Licence entries and full texts for Ollama (MIT), the Google Gemma 4 models (Apache-2.0), the Alibaba Qwen models (Apache-2.0) and ACE-Step (Apache-2.0) in Settings → Third-party licences.
No other change from 1.0.35; the plug-ins are the 1.0.35 builds.
v1.0.35
Fixes from a full-program audit, the corrected in-app guide, and the new user manual.
Fixed
- The in-app help pages open again. Settings → First Session Guide and DAW Compatibility handed the operating system a path inside the app's archive, which it cannot read, so the pages never opened in a packaged build. The full manual also lives at stemdesk.ai/manual.
- Analyze Track no longer classifies a track by its temporary file name. The role classifier keyword-matches file names and the capture's random name could contain "808"; a vocal could come back configured as a sub-bass. The app now classifies on the strip's real track name, or on spectral analysis alone.
- Cancel means cancel. Cancelling a queued stem split, or Auto-Mix while the output-folder dialog is open, used to run the job anyway and report success.
- A dropped connection during the AI model download no longer closes the app or leaves setup stuck until restart, and a failed download after consent asks again.
- Volume rides no longer click on bypass toggles, transport stop or when a new ride arrives, and a ride update can no longer free memory inside the audio callback.
- Pro Tools session build picks the stems' real sample rate for 32-bit float and RF64 WAVs; it used to fall back to 48 kHz.
- The render engine tracks plug-in latency after parameter changes (linear phase, oversampling, lookahead) so multi-pass renders stay sample-aligned.
- Duck depth uses the session's role for generically named stems; WhisperX vocal rides auto-detect the language.
- GPU detection survives one unparseable adapter line; a model download that ends early keeps its partial file for resume.
- Preset recall is atomic against the audio thread; bypass no longer snaps on oversized host blocks; the plug-in server never holds its connection lock across a blocking send.
- An older plug-in against a newer app is told why it was rejected (REINSTALL on the ANALYZE button) and backs off instead of reconnecting every two seconds.
- Windows: the plug-in manager helper refuses reparse points in its mailbox and verifies its write path before writing as SYSTEM.
- Security hygiene: plug-in server token no longer in the process list, saved report names validated, diagnostic reports redact API keys and tokens.
Changed
- EULA 1.4: Audio Units named, the Free-tier grant matches the shipped app, the macOS first-run component install is disclosed. Everyone re-accepts once.
- The Studio activation notice about iLok was wrong and is gone: no StemDesk plug-in needs an iLok.
- The update dialog links the full release notes when a note is long. DAW compatibility list refreshed. macOS first-run dependencies are now pinned to the tested set.
v1.0.34
Analyze Track hardening and installer hardening.
Changed
- Analyze Track captures up to 60 seconds (was 20). Longer captures give the analysis more to work with for tempo, sections and loudness range; the only cost is temporary disk space, deleted after analysis.
- Installer hardening. The packaged app ships with all five Electron fuses set strict: the signed binary ignores
NODE_OPTIONSand--inspect, cannot be run as a bare Node runtime, loads only from its own application archive, and refuses to start if that archive has been swapped or patched.
Fixed
- Analyze Track error messages were unreadable. The button cut every message to 14 characters and showed it for under a second. It now shows a short label (CONFIGURED, NO AUDIO, FAILED) with the full message in its tooltip, for 4 seconds on success and 10 on an error.
- Pressing ANALYZE without playing now says "No audio reached the strip — start playback, then press ANALYZE" instead of reporting a silent capture.
- A buffer-size change no longer aborts a capture with a false "sample rate changed"; only a real sample-rate change does.
- Old volume rides are no longer pushed to a strip after Analyze Track, and the strip's real stem association is no longer overwritten by the capture's temporary file name.
- A late reply from the app can no longer cancel a newer capture.
- The app accepts capture files only from its own captures folder.
- Two strips started in the same millisecond no longer write to the same file; the audio-thread tap guards against a zero-channel buffer; buffer overruns are counted and reported instead of silently splicing the capture.
- The First Session Guide no longer claims Analyze Track renames the strip.
v1.0.33
Plug-in fixes from two weeks of studio sessions and a fleet-wide audio audit, plus the first version of Analyze Track.
Added
- Analyze Track — configure the strip without bouncing. The Channel Strip has an ANALYZE button: press it, play the track, and the app analyzes the capture and pushes the complete configuration back: every processor and the chain order for the detected role.
- First Session Guide inside the app (Settings → First session guide).
Fixed
- Plug-in meters no longer freeze when the transport stops. In Pro Tools and other hosts that suspend plug-in processing at stop, nine plug-ins used to hold their last reading indefinitely.
- De-esser display. The gain-reduction edge no longer draws the wrong fill at the start and end of reduction spans.
- Vocal tuner. Reworked pitch surface from a Pro Tools listening session: steadier readout between notes, cleaner transitions.
- Saturator: Drive 0 is now a true bypass. Previously Drive 0 still applied a slight level boost to quiet material. Any Drive value above 0 sounds exactly as it did, verified bit-identical across all eight modes.
- A new automated audio-contract gate proves every plug-in's DSP passes silence, unity, NaN/Inf and DC checks on every build.
v1.0.32
A hardening release on top of the Audio Unit release.
Fixed
- The start-up component check no longer stalls a slow first launch; it reports each component as it is checked.
- A partially downloaded vocal-restoration model no longer disables vocal restoration permanently; damaged files are re-downloaded.
- Upgrading no longer leaves old plug-ins beside the new app. Installed plug-ins are compared to the shipped copies by content at every launch and replaced when they differ.
- Windows: the installer refuses to start an upgrade while a DAW has a StemDesk plug-in open, instead of updating the app and silently failing on the plug-ins.
Changed
- macOS: every plug-in bundle carries its own notarization ticket, so a StemDesk Audio Unit or VST3 copied out of the app validates on its own, offline.
v1.0.31
The Audio Unit release.
Added
- Audio Unit versions of all 17 plug-ins on macOS. They pass Apple's auval validator, are Developer ID signed and notarized, and connect to the StemDesk app from inside macOS's sandboxed AU host (verified in GarageBand, which uses the same hosting service as Logic Pro). External sidechain in AU requires a stereo main input.
- AU components install to
~/Library/Audio/Plug-Ins/Componentsand follow the same tier entitlement as the VST3s. - REAPER and FL Studio are listed as tested in-host on the compatibility page, alongside Pro Tools.
Fixed
- The AI critique can no longer lift a digitally silent stem.
- A fresh install on a machine that could not reach the vocal-restoration model host could not render at all; the render engine no longer depends on that download.
Earlier releases (v1.0.0 to v1.0.30) are summarized in the app's in-app changelog. STEMDESK™ is a trademark of Fulks, Inc. · © 2026 Fulks, Inc. · stemdesk.ai