Product Updates

Changelog. Every feature, fix, and improvement — tracked and transparent.

Release history
2026.09.24September 24, 20262 new1 improved

Stable Audio 2.5 in the Generator

  • NewA second music model. The Generate page now lets you choose Stable Audio 2.5 alongside ElevenLabs Music for each generation — instrumentals and sound design up to 190 seconds, delivered as WAV rather than compressed audio
  • NewStable Audio output carries no media carve-out: what you make with it can go into film, TV, radio and games. The exclusions in section 4 of the Terms apply to ElevenLabs output only
  • ImprovedStable Audio is priced flat per generation at any length, shown before you generate; ElevenLabs Music stays metered by length. Leave the model unset and everything — Studio, the Copilot, the API and MCP — keeps using ElevenLabs as before
2026.09.07September 7, 20265 new2 improved

Take Lanes, Swipe Comping, and the Rest of Tracking

  • NewComp by hand as well as by swipe. In the take lanes, double-click splits the section under the pointer, dragging a seam moves it (both sides trim together), and a click hands the section under it to that lane’s take. The lane’s name still means “the whole clip”
  • NewThe built-in instruments play the expression a take recorded. Pitch bend moves a sounding note in cents from the moment it was bent; aftertouch swells the note through its own gain stage. Both work in playback and in export, and a note without any expression sounds exactly as before
  • ImprovedThe Mac app says when two clocks disagreed. If the input and output were on different devices during a take, the toast that places it reports how many milliseconds of drift the engine corrected; if the engine could not keep up, it reports the frames it lost instead. Audio Engine settings show the same figures live while the input is open
  • NewTake lanes and swipe comping. Right-click a clip that has takes → Takes → Show take lanes, and each take opens as a lane under the clip. Drag across a lane and that stretch of the clip plays that take; click a lane’s name to use it everywhere. The clip above becomes the comp, with a short crossfade at every seam, and plays, exports and bounces like any other clip
  • NewMIDI replace mode and input quantise. Audio Engine → MIDI Recording chooses whether a take merges with the notes already there or clears the span it covers first, and snaps note starts to a grid as they are recorded — quarters to 32nds, triplets included, or off to keep the performance as played
  • NewThe Mac app measures its own latency. Calibrate now plays clicks from inside the engine on known frames and records them the way a take is recorded; what arrives late past the engine’s buffers — the converters, the room — is stored and taken off every take from then on
  • ImprovedRecording with the microphone on one device and the speakers on another no longer drifts. The engine watches how far the input runs ahead of or behind the output and nudges its read rate by a fraction of a percent until they agree; a shared clock stays bit-for-bit untouched
2026.09.06September 6, 20262 new1 improved

The Mac App Records Through Its Own Engine

  • NewIn the native Mac app, recording now happens inside the audio engine. The input is opened by the engine itself, so every recorded block is stamped with the exact frame of the song it played against — takes land on the sample, punch in and out land on the sample, and a loop wrap starts a new take file on the boundary. The app used to record through the web page on a separate clock, placing takes to within a frame of video rather than a frame of audio
  • NewMonitoring in the Mac app runs through the track. Each input feeds the armed track’s strip — fader, inserts, pan, sends — with the transport stopped or rolling, and the input meters in the track header read from the engine. A multi-input interface records one sample-aligned file, split per track when the take lands on the timeline
  • ImprovedThe engine subtracts its own buffer latency from every native take, so nothing needs calibrating there; the browser’s Calibrate button says so instead of running clicks the native app would not hear
2026.09.05September 5, 20265 new5 improved

Recording That Rolls With the Song

  • NewPress Record and the song plays. After the count-in the transport rolls from the playhead and the take starts on the same beat, so you record against the tracks you already have — the way a DAW works. Record used to capture the input while the timeline stood still, and if you pressed Play yourself the take landed where it stopped rather than where it started. Record again ends the take while the song keeps playing; Stop ends both
  • NewTakes land where they were played. A take is placed by the audio clock — the exact frame it began on — instead of by reading the playhead after the upload finished. Punch in and out land on their frames too, not on the next 50 ms check, and the input is opened before the punch point rather than at it
  • NewCycle recording makes takes. Set a loop and record: every pass over it becomes a take on the clip, the last one active, and the clip shows T3/5 so you can see it has them. Right-click the clip → Takes to switch. Loop recording used to need a manual stop for every pass
  • NewInput meters. Arm a track and its input level is live in the track header before you record, so you can set gain and see clipping. Nothing showed input level before
  • NewMeasure your real recording latency. Editor → Audio Engine → Calibrate plays five clicks through the speakers and measures when the microphone hears them; recordings are moved earlier by exactly that. The old compensation only knew the output half of the round trip and was off by default
  • ImprovedA single-input take is recorded clean. Recording onto one armed track went through the browser’s speech chain — echo cancellation, noise suppression and automatic gain — and was saved as compressed Opus. Every take is now uncompressed 24-bit WAV (16-bit or 32-bit float in Audio Engine settings) with the processing off, on the same path multi-input takes used
  • ImprovedMonitoring goes through the track. Input monitoring routes through the armed track’s fader, pan, effects and sends, so you hear what the take will sound like. It used to go straight to the speakers from a second audio device, dry, and fell silent during multi-input takes
  • ImprovedLong takes no longer fill the tab. Audio is written to storage as it arrives instead of being held in memory until you stop; the only limit is the upload size, and a take that reaches it says so
  • ImprovedMIDI recording keeps the sustain pedal, pitch bend and aftertouch. Pedalled chords hold for as long as the pedal was down; bend and pressure are stored on each note. Over a loop, passes merge into one clip the length of the loop. Audio and MIDI recorded in one pass share the same starting frame
  • ImprovedThe metronome follows the bars. With the click on, starting playback re-anchors it to the timeline so it clicks on the beats of the song rather than on a grid that started whenever the click button was pressed
2026.08.21August 21, 20263 new6 improved

Parts That Follow the Song, MIDI From Real Playing, and Instruments Worth Hearing

  • NewMIDI tracks can play through a real instrument. The voice picker on a MIDI track now lists plugin instruments under the built-in voices — a sampled Soundfont Player (piano, guitar, strings, brass, drums), Synth-101, Spectrum Modal and a DRM-16 drum sampler — and they load on first pick. Choose one and the track’s notes play through it, from the transport as well as from a keyboard
  • NewAudio-to-MIDI handles real playing. Convert to MIDI now runs Spotify’s Basic Pitch in your browser — polyphonic, so a fingerpicked guitar or a voice over piano comes through as the notes that were played, not “no pitched notes found”. Thirty seconds of audio takes well under a second once the model is loaded, and nothing is uploaded
  • NewA part generated to follow an existing clip actually follows it. Ask for a guitar over the piano take and the reference’s chord progression — every diminished and augmented passing chord, with its length — goes into the brief, and the result lands on the timeline where the reference starts instead of at zero. Generated MIDI parts compose over the reference’s detected chords bar by bar rather than over a key name
  • ImprovedAsking the agent for a full arrangement with separated stems now puts the tracks on the timeline. Stems arrive as one track each, aligned with the clip they were cut from; an AI Cover rebuild is added with its tempo and master chain. Both used to come back as a list of files and a suggestion to open a dialog, after the credits had been spent
  • ImprovedEffects on a track’s insert rack are audible during playback. The live mixer was never handed the effect builder the offline renderer uses, so every rack insert — and every plugin that was meant to run live — was drawn, saved, and silent until export
  • ImprovedClicking a “Your track is ready” notification opens your library on the Generated tab instead of a 404. Two older notification types (generation ready, model training done) pointed at pages that did not exist either
  • ImprovedAsking the agent to delete several tracks at once no longer leaves a card spinning forever under “Remove track “undefined””. Several short tool calls arriving in one chunk tripped a race in how the cards were indexed, which tore the whole reply down mid-stream
  • ImprovedThe clip analysis table reports the tempo and key a track was generated with, or tagged with in the library, ahead of what it reads off the waveform — and says which is which. The waveform estimate halves, doubles and wobbles on sparse voice-and-piano material
  • ImprovedThe agent has a route for adding a sung vocal line over an arrangement — lifting the vocal stem from your own recording, or generating one with lyrics over the arrangement as the reference — instead of declining
2026.08.19August 19, 20263 new8 improved

Chord Detection, and Lyrics That Reach the Model From Everywhere

  • NewRight-click an audio clip and choose Detect chords to find out what it is playing. You get the chord at each beat, the key they sit in, and each chord’s degree in that key — written to the clip’s chord track, with the symbols drawn along the top of the clip so you can see the changes against the waveform. The chord track had been in the piano roll for months and could only ever be typed into by hand: it could hold a progression but never hear one
  • NewIt runs in your browser and costs nothing — no credits, no model call, nothing uploaded. A three-minute track takes about a fifth of a second, which is the point: analysis you are charged for is analysis you don’t run on every take
  • NewAsk the agent what key or chords something is in and it can now listen instead of guessing. Once the chords are found, Render to MIDI turns them into playable voicings and Harmonize melody voices a line against them — so you can detect the harmony in a generated track and write a bass or pad that actually follows it
  • ImprovedThe agent was writing chords in the wrong place. Its tool said chord times were in seconds while the chord track has always drawn them in beats, so at 120 BPM every chord it wrote landed at half its intended position and the last ones fell off the end of the clip entirely
  • ImprovedThe chord track shows each chord’s degree — I, vi, bVII, V7 — under its name, instead of respelling the notes that are already in the symbol above it
  • ImprovedAsk the copilot for an instrumental and get one. The tool it calls had no vocals switch — only the Generate page did — so "no vocals" could only be phrased as a wish inside the prompt, and the model decided. It is now the same explicit setting the page has
  • ImprovedTell the copilot a tempo or a key and it passes them on as a tempo and a key. They previously reached the model only if it happened to restate them in the brief it wrote, so asking for 85 BPM in A minor was a suggestion rather than an instruction
  • ImprovedGenerating over the API honours the BPM, key and genre it accepts. All three were listed as settings, sent on every request, and read by nothing on the other end — so a track requested at 85 BPM in A minor was composed at neither. Lyrics and an instrumental switch are available there now too
  • ImprovedA track generated from a reference now carries the tempo and key measured off that reference, so it arrives in the Studio with a tempo instead of a blank
  • ImprovedAPI generations appear once in your history. Every one was recorded twice — the finished track, plus a second copy stuck at "pending" that nothing ever completed
  • ImprovedA prompt that already names a key is no longer contradicted by the key picker. "In the key of Am" plus a picker on C sent both and let the model choose
2026.08.18August 18, 20265 new8 improved

Vocals That Sing Your Words, Live MIDI Performance & Real Plugins

  • NewAsk for a song with words and get the words. Generate now has a Vocals control — With vocals or Instrumental, plus a lyrics box — and sends both to the model. The page had never sent either, so anything you typed arrived as a style brief: ask for a song that says a particular line and you got music vaguely about it, with words the model invented or none at all. The help text told you to supply lyrics, describing a field that did not exist
  • NewPlay a connected MIDI controller and hear the selected track — its built-in instrument, or an instrument plugin sitting in its rack, picked from where the track already points rather than from a separate setting
  • NewThe plugin panel shows plugins you actually have. It used to list Serum, Massive and Kontakt from a hardcoded array, behind a Scan button that was a two-second timer reporting the length of that same array — none of them loadable. It now reports real state, and where the answer is no it says which format and why, so an empty folder no longer looks the same as a build that cannot host anything
  • NewWeb Audio Module plugins load and play in the browser, live and in a bounce, with a panel for the parameters the plugin exposes
  • NewA bug report you start still reaches us if you close the tab before it finishes sending
  • ImprovedThe waveform under a track is now that track’s waveform. When audio failed to decode, the player drew a procedural bar pattern that looked exactly like a real one, so a file that never loaded was indistinguishable from one that did — and the failure was swallowed, leaving nothing to find
  • ImprovedGenerated tracks come back the length you asked for. The duration was being sent under a name the music model does not read, so it was quietly ignored
  • ImprovedA malformed project now opens wrong instead of taking the Studio down with it. One stray empty clip or effect — from a DAW import, a cloud restore, or a project file you loaded — reached the renderer and the whole screen went, so a slightly odd project read as an unopenable one
  • ImprovedOne MIDI connection for the whole app. Every part of it asked the browser separately, so the device picker governed only the dialog it lived in, recording stayed subscribed to every device on the system, and panels that never asked were silently deaf. The permission prompt could also appear more than once
  • ImprovedA recording that fails to upload no longer lands on the timeline as a clip pointing at nothing
  • ImprovedLive MIDI holds up to what controllers actually send: a long identity reply on connect, and timing bytes interleaved between a message’s data bytes — which the spec allows, and which used to turn one note into a different one
  • ImprovedCrash and error reports reach us again. They had been failing silently in the desktop app and on the site, so problems you hit left no record for anyone to work from
  • ImprovedEffects no longer stay broken after one bad sample of audio. A single corrupt value inside a filter would keep circulating through it, and the track stayed silent until you happened to move the playhead
2026.08.16August 16, 20266 new17 improved

Hands-On Agent Replies, Audio to MIDI & Fourteen Instruments

  • NewThe agent hands you the control, not just the result — when it loads a plugin, moves a fader or changes the tempo, its reply carries the real knobs underneath, so you can take the value somewhere else without typing another sentence. Every one has a Revert
  • NewConvert an audio clip to MIDI — a hummed line, a recorded riff or a stem becomes notes in the piano roll, with pitch and timing detected from the audio rather than quantised onto a grid
  • NewFourteen built-in instruments, up from four, and the agent can look them up by name — so "make the bass a Rhodes" picks the Rhodes instead of the nearest of four things
  • NewThe Studio has a real layout system instead of stacked full-width rows, and the left nav collapses to an icon rail when you want the screen for the arrangement
  • NewBar numbers on the timeline ruler, and double-clicking a clip opens its editor
  • NewNew agent abilities: loop a clip, convert a clip to MIDI, and build an FX send bus — which adopts the tracks you were talking about when you do not name any
  • ImprovedGenerating audio no longer freezes the page. The analyzer that ran over each new track was doing the slow kind of Fourier transform on the whole file, which is fine for a drum hit and takes hours for a song; a three-minute track now analyses in about a fifth of a second instead of locking the tab
  • ImprovedYour BPM and key now actually reach the music model. They were being folded into an "energy" number and a prose description and then guessed back out, so 85 BPM in A minor could arrive as 120 in C
  • ImprovedGenerating charged you twice — once in the browser and once on the server, at two different amounts, neither of which was the number on the button. The server charge is the real one and the browser no longer writes its own
  • ImprovedTempo detection stops reporting half the tempo on anything with a backbeat. It was counting kick and snare as one pulse; it now listens across frequency bands and measures the spacing rather than voting on it
  • ImprovedMIDI files that imported as "empty" import properly — the reader was assuming a fixed header size and giving up at the first unrecognised chunk, both of which the format explicitly allows
  • ImprovedImported and generated clips are the length of the audio in them. When the quick duration check failed the clip silently became a 4- or 30-second constant, so a three-minute song landed as a four-second clip and had to be dragged back out by hand
  • ImprovedA new project can start at any tempo you type, instead of the handful on a dropdown
  • ImprovedWhen a track goes quiet mid-playback because its audio could not be scheduled, it now says so and names the track. The playhead kept moving and nothing explained the silence
  • ImprovedLong sessions stop going silent — audio links are re-signed while you work rather than expiring under you
  • ImprovedCover art shows up again in the Generated tab; the thumbnails broke when audio storage went private, and they belong in a public bucket rather than behind an hour-long signed link
  • ImprovedDeclining the microphone prompt no longer takes the Audio Editor down with it
  • ImprovedBass notes no longer clip at full velocity, and every instrument now has a defined tail rather than cutting off or ringing on
  • ImprovedThe reverb is finished properly and wet/dry parallel processing is one click. Distortion modes are named after what they sound like instead of numbered
  • ImprovedThe agent and the Bus menu agree on what is allowed to feed a bus, so a routing the agent offers is one the mixer will accept
  • ImprovedThe agent stops describing tools it does not have: its list is generated from the same help you can read
  • ImprovedAsk the agent about the desktop app, share links or the artwork on your generated tracks and it can answer. All three had shipped without ever being written into the help it reads, so it would tell you they did not exist — and the same gap reached the in-app help search and every connected MCP client
  • ImprovedSmaller fixes: every effect exposes its parameters in the UI, the empty-timeline card stays on screen on a narrow window, and repeated rate-limit warnings collapse into one
2026.08.15August 15, 20268 new16 improved

Desktop Apps, 47 Stock Effects & Private-by-Default Audio

  • NewSoundGen for macOS and Windows — download the native desktop app at /download. The mac build is signed and notarized by Apple, so it opens without a Gatekeeper warning
  • NewA wave of new stock effects: tilt EQ, graphic EQ, vibrato, doubler, trance gate, tape wobble, frequency shifter, gated reverb, shimmer reverb, beat repeat, expander, glue compressor, opto compressor, mono maker, mid/side, and stereo imager — with honest fallbacks where the built-in Web Audio nodes cannot express the DSP
  • NewMetering inserts — a spectrum analyzer, a phase/correlation meter and a chromatic tuner that drop into any chain like any other device. The tuner takes an A4 reference, so a room tuned to 442 or 415 reads correctly instead of permanently sharp
  • NewReport a bug or request a feature without leaving the app — /bug and /feature in the agent write the report up, attach the page, build and recent errors (with credentials stripped out before anything leaves your browser), and file it. Ask later and it tells you where each of your reports stands
  • NewPublic project player at /play/:id — share a whole project, not just a single track
  • NewPer-track visibility — audio is private by default and only reachable publicly after you explicitly publish it, with Public/Private toggles across My Sounds
  • NewClip fade curves and MIDI note transforms in the piano roll
  • NewThe agent can now search product help, and publish or unpublish your sounds and generations on request
  • ImprovedAudio storage is now private end to end — playback goes through a signed-URL broker that authorizes each file against its owner or its published status, so an old copied URL no longer grants permanent access
  • ImprovedAgent replies no longer cut off mid-sentence — the streaming parser was dropping the final frame, and the output budget was raised from 2k to 32k tokens
  • ImprovedDeep stem separation is priced for the work it actually does: it runs a 6-stem pass plus a second pass splitting the drum bus, so it now costs more than a standard single-pass separation instead of the same flat rate
  • ImprovedLong generations got much cheaper — the markup tapers with length instead of sitting flat, because the fixed cost of a request is the same whether you ask for 15 seconds or five minutes. A five-minute track is now 125 credits instead of 200
  • ImprovedSeparation prices are derived from the providers' published per-run costs rather than an estimate: a standard job is 6 credits (was 10) and a deep one 17 (was 25). A deep job whose drum-split pass cannot run is billed for the pass that did, instead of charging the full cascade price for stems you did not receive
  • ImprovedThe agent quotes real prices. The figures behind its pricing answers were typed in by hand and had gone stale — it was quoting one flat number for both separation modes — so they now come from the same pricing code that charges you
  • ImprovedThe agent can see the whole effect catalogue again. The size cap on tool results was cutting the list short, leaving the fifteen newest effects invisible to it and the truncated response unparseable
  • ImprovedFixed a crash that dropped the entire app to a blank "Loading…" screen — closing a realtime channel re-entered its own teardown until the stack ran out, which is what took the studio down after an agent turn or a project import
  • ImprovedDynamic EQ no longer comes out 12 dB louder than it went in — its four bands were summing in parallel rather than sitting in series
  • ImprovedBeat repeat and expander load properly; a stray backtick was cutting their worklet source short
  • ImprovedStem separation model pins refreshed after the old versions stopped resolving
  • ImprovedRecent projects no longer shows an unsaved draft alongside the saved copy of the same project
  • ImprovedSmaller studio fixes: the export dialog says where MIDI export lives, the MIDI editor tells you whether you need a track or a clip, and scene names in Session view stop truncating to "Sce…"
  • ImprovedPricing tiers, their feature bullets and every quoted product number are now generated from the code that produces them, guarded by a test that fails on drift
  • ImprovedStale lazy-loaded page chunks recover instead of erroring after a deploy
  • ImprovedPublic beta is stated plainly in the page title, social preview, footer and pricing
2026.08.14August 14, 20267 new6 improved

DAW Project Import, Mixer Busses & MIDI Recording

  • NewImport projects from other DAWs — Ableton Live, REAPER and DAWproject import arrangement, clips and timing in full; Logic Pro, GarageBand and FL Studio recover tempo, key, time signature and track names
  • NewMixer bus and send tier — route tracks to busses and set up sends, with plugin delay compensation across the graph
  • NewMIDI recording, plus a piano roll you can edit by hand
  • NewAutomation write modes with fader automation driven through AudioParams
  • NewMaster EQ and stereo master metering
  • NewTake comping — record multiple takes and assemble the keeper
  • NewGenerate audio directly into a marquee selection on the timeline
  • ImprovedThe agent got full Studio control and a measure/act/verify loop, so it checks its own work instead of narrating it
  • ImprovedOffline bounce now agrees with live playback — MIDI renders offline, automation shares the same maths, and the metronome actually sounds
  • ImprovedThe audio editor writes its edits back to the clip instead of discarding them
  • ImprovedCredit delivery fixed end to end: purchases and welcome bonuses that were silently failing now land, and the contact form reaches a real inbox
  • ImprovedPassword reset routing and reset-link expiry fixed
  • ImprovedThe DAW importer is hardened against malformed and hostile project files
2026.08.04August 4, 20264 improved

Credit Reliability

  • ImprovedCredit deduction flow corrected so charges land exactly once against the canonical credit type
  • ImprovedInsufficient-credit responses are no longer misread as a generic failure, so you get told what actually went wrong
  • ImprovedSafety rejections from the model are surfaced properly and retried where retrying is valid
  • ImprovedFixed a double-generate bug that could charge twice for one request
2026.07.28July 28, 20261 new1 improved

Reference-Track Generation & Changelog Refresh

  • NewGenerate Similar tool — upload a reference track, auto-analyze BPM/key/mood/instrumentation, and generate a new song in the same style
  • ImprovedChangelog is now a dedicated /changelog landing page — no more first-visit popup dialog
2026.07.25July 25, 20263 new4 improved

Public Sharing, Downloads & Track Editing

  • NewPublic share links — flip any generated track to Public and copy a soundgen.ai/s/<id> link that plays for anyone, no sign-in required
  • NewEdit generated tracks — rename and update descriptions from the Generated tab, changes sync to the public share page
  • NewOne-click MP3 download on every generated track in My Sounds
  • ImprovedShare links always resolve to the soundgen.ai domain instead of preview URLs
  • ImprovedAnonymous read access on public generations via updated RLS grants
  • ImprovedDownloads now use the correct .mp3 extension instead of .mpeg
  • ImprovedGenerated card action row wraps cleanly on narrow layouts
2026.07.22July 22, 20264 new4 improved

AI Music Pipeline: Background Jobs, Cover Art & My Sounds

  • NewGenerated tab on My Sounds — every AI-produced track lands here automatically with cover art, waveform, and playback
  • NewAuto-generated AI cover art on every track using Gemini 3.1 Flash Image, bundled free with music generation
  • NewPending Generations bar above the composer — expandable list with per-job elapsed time, live pipeline step, and cancel button
  • NewBackground music generation — keep chatting with the agent while a track renders; playable card appears when ready
  • ImprovedServer-side persistence in the elevenlabs-music edge function: upload → ai_generations row → credit charge → cover art all complete before the client can drop the connection
  • ImprovedAgent forces a real generateMusic tool call on "generate a full track" requests instead of narrating fake progress
  • ImprovedPasted lyrics are auto-detected (verse/chorus/bridge markers) and passed as the lyrics param so ElevenLabs actually sings them
  • ImprovedAI sidebar now exposes the full agent workspace — conversation list panel, rename/delete, new-conversation button, attachments, slash commands, and stop-and-steer
2026.07.15July 15, 20265 improved

Credits, Reliability & Playback Hardening

  • ImprovedServer-side credit deduction on every paid generation — credits always debit exactly once, independent of client-side flow
  • ImprovedWelcome-gift copy correctly shows the 100 credit bonus instead of echoing current balance
  • ImproveduseAudio.play skips dead blob URLs cleanly instead of throwing NotSupportedError
  • ImproveduseRecentlyPlayed filters out sounds whose blob URLs no longer resolve
  • ImprovedRemoved "SoundGen" wordmark from the footer logo to match the header treatment
2026.05.28May 28, 20261 new6 improved

Modular Studio Refactor, Stability & UI Polish

  • NewTheme-aware SoundGen logo — light and dark variants swap in across the marketing header, footer, and sidebar
  • ImprovedProduction builds ship with a Supabase URL + anon-key fallback so the app no longer fails to boot when env vars are missing from the build pipeline
  • ImprovedRemoved the floating Team Chat widget that was overlapping the AI Agent button on Studio pages
  • ImprovedStudio refactor: Timeline, Discover, Arrangement, AI Features panel, Audio Export, Credits, and Keyboard Shortcuts split into focused modules for faster iteration and smaller bundles
  • ImprovedAudio engine refactor: effect factories split (basic / modulation / processing / dynamic EQ) and worklet processors extracted into per-file modules (pitch-shift, spectral-gate, limiter, mono-check, gain)
  • ImprovedEdge functions (ai-audio, cred-api, mcp-server) broken into per-tool / per-handler files for clearer ownership and easier review
  • ImprovedAudio editor: extracted `useEditorWaveform` hook and lighter waveform render path in the advanced editor
2026.05.27May 27, 20263 new5 improved

Spectral Audio Engine: Pitch, Noise & Vocal Isolation

  • NewVocal isolation — one-click vocal extraction using the Demucs stem engine
  • NewSpectral-gate noise removal — FFT worklet with minimum-statistics noise tracking that preserves held notes and vibrato
  • NewAdaptive de-click with frozen last-good trajectory + clamp to repair spikes without runaway artifacts
  • ImprovedPitch shift is now true stereo — per-channel phase-vocoder state (no more silenced right channel)
  • ImprovedVocal tuning and noise removal both run through AudioWorklet with safe Web Audio fallbacks
  • ImprovedEffect-chain rebuilds no longer leak panner nodes — wrapped multi-node effects override both connect and disconnect
  • ImprovedDemo-mode toast warning surfaced when REPLICATE_API_KEY is missing for separation jobs
  • ImprovedStem separation guards against blob:/data: URLs and probes real clip duration before submitting
2026.02.26February 26, 20265 new2 improved

AI Agent Platform, Mac Desktop & Theme System

  • NewAI Agent — dedicated /agent workspace with toolkit, auto-run + undo, and saved multi-conversation history
  • NewGlobal right-sidebar Copilot — chat available from any page, retains project context
  • NewTauri 2 Mac desktop build — `bun tauri:dev` for local, `bun tauri:build` for a real .app
  • NewLight / Dark mode toggle in the header with full design-token coverage
  • NewVoice Cloning Dialog: file upload to Supabase storage + URL input + multi-step wizard
  • ImprovedStudio AI panel is the single entry point for voice cloning, stem separation, and all AI features
  • ImprovedLight-mode contrast pass: text-white → text-foreground convention everywhere except colored fills
2026.02.25February 25, 20268 new4 improved

AI Copilot, Generation Pillar & Audio Wiring

  • NewAI Copilot in Studio — chat with action commands: setBPM, setMasterVolume, addTrack, muteTrack, soloTrack, setKey
  • NewText → Music via ElevenLabs `/v1/music` with mode toggle (Music / Loop / SFX) and inline preview
  • NewText → Loop and Text → SFX (ElevenLabs `/v1/sound-generation`) with inline players
  • NewChord detection, melody continuation, and reference/style matching from prompts and URLs
  • NewStem separation via Replicate / Demucs — each returned stem becomes its own timeline track
  • NewAI search on Discover — natural-language BPM, genre, and tag filtering
  • NewReal-time pitch shift effect powered by an AudioWorklet phase vocoder (±24 st)
  • NewVocal tuning effect — worklet with high-shelf + compressor fallback
  • ImprovedMute / Solo / Effects now drive the real Web Audio graph from `updateTrack`
  • Improved3-band EQ (low-shelf / peaking / high-shelf) on every track, live-updated from the Channel Strip
  • ImprovedMaster volume + master effect chain inserted between masterGain and limiter
  • ImprovedBuilt-in synth routes through the project masterGain so it lands in the mix
2026.02.21February 21, 202617 new

Collaboration, Import/Export & Platform Infrastructure

  • NewStem export — export individual tracks as separate files
  • NewMIDI import/export — import and export standard MIDI files
  • NewBatch export with configurable naming patterns
  • NewExport presets — save and load format/quality configurations
  • NewShare link with embedded playback for public projects
  • NewEmbeddable player widget for websites and blogs
  • NewExport to streaming platforms (SoundCloud, YouTube, Spotify, Bandcamp)
  • NewVersion branching — create, switch, and merge project branches
  • NewTrack locking — lock tracks so only one person can edit
  • NewTimeline comments — pin comments to specific time positions
  • NewActivity feed — see who changed what and when
  • NewMerge conflict resolution for simultaneous edits
  • NewOffline mode — work without internet, sync when reconnected
  • NewPlugin marketplace — browse and install third-party audio plugins
  • NewLocalization (i18n) — multi-language UI (EN, ES, FR, DE, JA, ZH, KO, PT)
  • NewPerformance profiler — live CPU, memory, FPS, and latency monitoring
  • NewLanguage selector in settings and studio
2026.02.20February 20, 202611 new3 improved

Audio Engine & Professional Rendering

  • NewWeb Audio worklet processing for lower latency DSP
  • NewSample rate selection — 44.1k, 48k, 96k project sample rates
  • NewBit depth options — 16-bit, 24-bit, 32-bit float rendering
  • NewOffline bounce — faster-than-realtime export
  • NewConvolution reverb — load impulse responses for realistic spaces
  • NewParametric EQ visualizer with interactive frequency curve
  • NewCompressor gain reduction meter
  • NewBrickwall limiter on master bus to prevent clipping
  • NewTPDF dithering on export when downsampling to 16-bit
  • NewMono compatibility check — one-click mono fold-down
  • NewAudio buffer size control (128–2048 samples)
  • ImprovedWet/dry mix per effect — parallel processing control
  • ImprovedReal-time effects bypass — A/B toggle for entire chain
  • ImprovedFLAC, OGG, AIFF format support for import and export
2026.02.19February 19, 202610 new5 improved

UI/UX Polish & Studio Enhancements

  • NewDark/light theme toggle in the header
  • NewCustomizable accent colors — 8 selectable color themes
  • NewCommand palette — ⌘K search for any action or navigation
  • NewMinimap timeline — birds-eye view for quick navigation
  • NewUndo history panel — visual list of all actions
  • NewStatus bar — sample rate, bit depth, duration, CPU usage
  • NewOnboarding tour — guided walkthrough for new users
  • NewRight-click context menus on clips
  • NewLoading skeletons — content-shaped placeholders
  • NewFull-screen mode for maximum workspace
  • ImprovedResizable panels — drag to resize timeline/mixer/effects
  • ImprovedDrag-and-drop file import onto timeline
  • ImprovedResponsive Studio optimized for tablet and touch
  • ImprovedARIA labels and focus management for accessibility
  • ImprovedConsistent tooltips with keyboard shortcut hints
2026.02.18February 18, 20267 new3 improved

Core Platform Features

  • NewProject templates for quick start
  • NewEffect presets save & load
  • NewGeneration prompt suggestions
  • NewPWA support — install SoundGen on mobile
  • NewChangelog / What's New dialog
  • NewVertical track height zoom
  • NewCount-in metronome before recording
  • ImprovedEffects can now be reordered via drag buttons
  • ImprovedProject duplication from dashboard
  • ImprovedSound sort options: newest, popular, downloads