Manuwave Studio

Make complete songs — music, lyrics and cover art — with AI that runs entirely on your own computer. No subscription, no cloud, no upload. Describe an idea (or roll the dice), and Manuwave Studio writes a production brief, renders the track with ACE-Step 1.5, paints a cover, and lets you master and export it. Or make the music yourself: synthesizers, a drum machine and a MIDI keyboard, right in the browser — or a loop machine whose generators play on their own while you turn the knobs, and that hands your jam or your voice to the music AI to finish the song.

Listen to music made with it on Fahrstuhlmusik — the project's hobby showcase, where you can also share your own tracks.

The interface speaks English and German — it follows your browser language, and the button in the top bar switches it. The dice writes its ideas in the language you use.

What changed in each version: CHANGELOG.md.

What it does

🎼 Idea → song	One sentence is enough. A local language model turns it into a brief: sound tags, tempo, key, structure and original lyrics. Already run Ollama yourself? Pick one of your own models under ⚙️ — it is used directly, nothing is downloaded twice.
🧬 From your songs	Upload 1–3 songs: your computer measures tempo, key, chords, groove, instrumentation, energy and structure, the AI writes a new song with their character — no melody or lyrics taken over.
🎲 Dice	Picks from 114 styles in 11 groups — from melodic techno to bossa nova, metal, film score — and writes a detailed idea, then renders right away.
🎶 Music AI	ACE-Step 1.5 via acestep.cpp renders instruments, vocals and mix in one pass.
🎨 Cover art	FLUX.2 [klein] 4B (macOS) or Z-Image Turbo (Windows/Linux) paints a cover that fits the mood.
🔎 Quality check	A listening-judgement model (Meta Audiobox Aesthetics) rates every finished take on your computer. Weak takes are rendered again automatically (up to twice, your choice) and the best one is kept; discarded takes stay in renders/_verworfen so you can check the verdict.
⏱ Long tracks	Up to 10 minutes. Beyond 4 minutes the music AI gets unstable, so the studio renders a main part and then extends it section by section, continuing from the last 30 seconds and crossfading the seam.
🎚 Mastering	Quick knobs plus a parametric 8-band equalizer over the track's spectrum, a dynamic EQ against AI harshness, a de-esser and a glue compressor, mono bass, stereo width, fades and auto-loudness for streaming (−14 LUFS), −16 LUFS or club (−9 LUFS). Save your own presets, keep favourites, let “My sound” learn from your tracks, and give every new track your sound automatically — what you hear is what you export.
⬇️ Export	MP3, FLAC, ALAC, WAV, AIFF, WavPack, TTA — with cover, tags and lyrics embedded.
📦 Release package	One ZIP with the mastered file, a 3000 × 3000 cover, upload metadata with AI-disclosure checkboxes, lyrics, a proof-of-origin PDF and a licence template (these two documents are in German for now).
🎹 Make your own	A third tab with nine basic synthesizers, 18 recreations of classic instruments (TB-303, Minimoog, Juno-60, DX7, Rhodes, B3 …) and a drum machine — every knob included — plus a piano roll, patterns, a song arranger and a loop recorder. Play it with a USB/MIDI keyboard or the computer keyboard. Create track renders your piece into the library, where it is mastered, exported and published like any other track.
🔁 Loop machine	A fourth tab of generators that run on their own — beat, bass, chords, arpeggio, melody, a free oscillator, your own bars and a mic loop. Start one and it comes in on the next bar; turn frequency, resonance, highs, echo, reverb and more while it plays. Take over the rest builds a band around your bars or your voice; the autopilot arranges.
🤖 AI takes over	From the loop machine: the music AI continues your jam (your bars are the start of the song) or plays a band around a loop you sang — with your real voice kept on top. The result is an AI track; its note says which part is yours.
🎼 MIDI image	Every AI song is turned into notes on your computer — drums, bass, chords and melody, on the beat. Open it on the instruments with one click (Play) and change anything, or download it as a MIDI file for any DAW. Optional; runs only while your computer is otherwise idle.
🔏 Honest by design	Every file is labelled in its tags — “AI-generated”, or “self-played” for what you made in Make your own (then only the cover counts as AI); an inaudible watermark (macOS/Linux) proves where it came from.

Make your own

🎛 Synthesizers	Analog poly and mono (three oscillators plus sub, ladder filter with a high-pass in front, two envelopes, two LFOs, unison, glide), supersaw, wavetable (six sets: analog, pulse width, sync, wavefolder, vowels, glass), string machine, 4-operator FM with eight algorithms, electric piano, tonewheel organ with nine drawbars, percussion and rotary speaker, and a plucked string (Karplus-Strong). Arpeggiator on every synth.
🏛 Classics	18 recreations, each with its own factory sounds (110 in all): TB-303 (accent; slide when notes overlap, like the original), Minimoog, Juno-60 (with its chorus), Jupiter-8, Prophet-5, OB-Xa, SH-101, MS-20, ARP Odyssey, CS-80, Taurus bass pedals, PPG Wave, DX7, M1 house organ, Rhodes, Wurlitzer, B3 and Clavinet. Save your own sounds.
🥁 Drum machine	16 pads, all synthesized (no samples) — TR-808, TR-909, TR-606, CR-78, 80s, trap, techno and lo-fi kits, each pad with level, pan, tune, decay, tone and snap. Step grid with accents.
🎛 Knobs	Click anywhere on a knob's arc and the value jumps there — nothing follows the mouse afterwards. For fine moves hold Shift and drag vertically, or use the mouse wheel; Alt-click restores the default. Knobs grow under the mouse.
🖥 Full width	The tab uses the whole monitor: from 1800 px tracks, editor and song sit side by side, and the on-screen keyboard shows as many octaves as fit — five (61 keys, like a common 61-key keyboard) on a wide screen.
🎼 Patterns & song	Eight patterns (1–16 bars) per project, piano roll and step grid, swing, quantize, undo. Chain patterns into a song. Up to 16 tracks with volume, pan, EQ, chorus and sends to a shared reverb and tempo-synced delay.
🎹 MIDI	Any class-compliant USB/MIDI keyboard: velocity, pitch bend, mod wheel, sustain pedal, volume fader, transport buttons in DAW mode, and MIDI learn for every knob. Web MIDI needs Chrome or Edge; without it (Safari) the computer keyboard plays (A W S E D … = C … C, Z/X octave).
🎼 AI songs on the instruments	Play in the library (or Song from the library here) opens a song's MIDI image as a project: drums on the matching kit, Minimoog bass, Prophet chords, Minimoog melody, in parts of 16 bars, at the song's measured tempo (the music AI rarely hits the requested BPM exactly).
⏺ Recording

The classics are our own recreations — no samples, no code from the originals. Their names only say which instrument inspired them; all trademarks belong to their owners, and there is no affiliation.

The sound runs in one AudioWorklet (band-limited oscillators, zero-delay-feedback filters), so timing stays tight even when the tab is in the background, and the offline render is sample-identical to what you hear.

Play on an AI song in the library: its MIDI image opens on the instruments — drums, bass, chords and melody at the measured tempo.

Loop machine

▶︎ Generators	Up to twelve, each with its own loop length (1–8 bars). Press ▶︎ and it comes in on the next bar, ■ and it stops at the end of the bar (double-click: right now); space starts or stops all. 🥁 Beat — eleven styles, density, variation, human feel, fills. 🎸 Bass — eight-note pulse, long roots, house offbeat, disco octaves, acid (303), funk, walking. 🎹 Chords — pad, halves, stabs, offbeat, eighths, with 7ths and 9ths, smoothly voiced. ✨ Arpeggio and 🎶 Melody (a motif that returns and develops). Every generator plays any sound of Make your own.
〰️ Oscillator	A free tone generator: sine, triangle, saw, pulse, supersaw, white/pink/brown noise; frequency 20 Hz–8 kHz (free, or snapping to the key), fine tune, unison, sub, FM, drift, glide, an LFO on pitch/volume/pulse width/pan/FM, and a gate in time for pulsing tones.
🎼 Your bars	Play a few bars after a one-bar count-in (computer keyboard, MIDI or the keys on screen). The machine plays them and spins them on — life of its own sets how far it strays; every fourth pass returns to your original.
🎤 Mic loop	Sing or play into the microphone, in time, with latency compensation; overdub up to eight layers, undo, stutter, reverse. To notes hears out your melody and gives it to a synth that carries it on.
🎛 Live sound	On every generator: filter frequency, resonance, high-pass, a tempo-synced wobble, lows/mids/highs, drive, ping-pong echo in time, feedback, reverb, pan, volume, mute/solo. ⤢ (or Sound) opens all settings of a generator large, across the whole width — generator, orb & global, live sound, recording, MIDI out and every knob of the instrument, side by side over the other generators; the keyboard stays free and plays that generator (Esc closes). Master: DJ filter, reverb length, level, limiter at −1.5 dBFS. MIDI learn puts any knob on a controller knob.
👁 See what you hear	When the machine changes something — the wobble sweeping the filter, the autopilot closing the filter for a breakdown, the oscillator snapping to the key, gliding, drifting, its LFO or gate — the knob turns with it, and the number shows the value that is sounding right now; your own setting stays as a small dot on the ring. Knobs that have no effect at the moment (oscillator 2 at level 0, pulse width on a saw wave, LFO without a destination …) are dimmed and say why; turning one switches the missing part on.

Requirements

	macOS	Windows / Linux
Hardware	Apple silicon, M1 or newer, 16 GB RAM recommended (8 GB works with smaller models)	NVIDIA GPU, RTX 20xx or newer, 8 GB VRAM; driver ≥ 550
System	macOS 14 or newer	Windows 10/11 · Ubuntu/Debian x86_64
Disk	about 25 GB (+1 GB for the optional quality check, +1.5 GB for the optional MIDI image, +2.6 GB — 5.3 GB with --xl — for the optional loop machine AI)	about 25 GB (+1 GB quality check, +1.5 GB MIDI image, +2.6/5.3 GB loop machine AI, all optional)
Status	✅ tested	⚠️ experimental — built and checked in CI, not yet run on real NVIDIA hardware by the maintainer

Install

1. Download the latest release ZIP (or git clone this repository) and unpack it.
2. Run the installer for your system:
   • macOS: double-click install/macos/Install Manuwave Studio.command (Homebrew is installed if missing — macOS asks for your password once).
   • Windows: double-click install\windows\Install Manuwave Studio.bat.
   • Linux: ./install/linux/install.sh
3. The installer downloads everything from the original sources (about 15–20 GB of models), builds what needs building, and creates a Manuwave Studio launcher (Desktop on macOS/Windows, app menu on Linux). Interrupted? Just run it again — downloads resume and finished steps are skipped.

Everything lives in ~/ManuwaveStudio (%USERPROFILE%\ManuwaveStudio on Windows). Your music is stored there under renders/ and in your browser's library.

Option	Effect
--lm 0.6B | 1.7B | 4B	size of the music AI's planning model (default: chosen by RAM)
--xl	the larger, better-sounding XL model (+3 GB)
--skip-writer, --skip-image	skip the language model or the cover model
--skip-qa	skip the quality check (PyTorch + rating model, about 1 GB; runs on the CPU)
--skip-midi	skip the MIDI image (Demucs + Basic Pitch, about 1.5 GB)
--skip-loopai	skip the loop machine AI (the base model for band for your voice, 2.6 GB or 5.3 GB with --xl); continuing a jam still works
--no-start	don't start the app at the end
--home DIR	install somewhere else (must be inside your home folder)

Windows: -Lm, -Xl, -SkipWriter, -SkipImage, -SkipQa, -SkipMidi, -SkipLoopAi, -NoStart.

Update

From 3.23 on, Manuwave Studio updates itself with one click:

• When a new release is out, ⬆ Update x.y.z appears at the top right. It opens the release notes and an Update now button.
• One click downloads the release from GitHub, checks it (it must compile and start on a spare port first), keeps the previous version in app.prev and restarts — the page reloads by itself. Your music, library, settings and pairing stay as they are. While a job is running, the update waits.
• The app asks GitHub for the latest release at most every six hours. You can switch that off under ⚙️ → Updates, and check by hand there.
• If a release needs new components (models, tools), the notice says so and you run the installer again instead: close the Manuwave Studio window first, then download the new release ZIP and run the installer with the same options. Finished steps are skipped, so models are not downloaded again.

Versions before 3.23 don't have the updater yet. Update them once with the installer as described above.

Git clones made before 2026-10-21: the repository history was rewritten on that day, so git pull refuses to merge. Clone the repository again (or git fetch origin && git reset --hard origin/main if you changed nothing).

If something doesn't work

• Keep the start window open. Manuwave Studio (the window that opens when you start it) runs the app and the bridge, and starts either of them again if it stops. Closing it stops everything.
• 🔴 App not responding at the top right means the app is not running: open the start window again. The page carries on by itself once it's back.
• Logs are in ~/ManuwaveStudio (%USERPROFILE%\ManuwaveStudio on Windows): app.log, bridge.log, and app.restart.log / bridge.restart.log after a restart. They help when you report a problem.
• Windows installs before 3.24.1 could not run in Windows PowerShell 5.1. Download 3.24.1 and run the installer again — finished steps are skipped.

How it works

Browser (UI, library)  ⇄  app server (127.0.0.1:8104)  ⇄  bridge  ─┬─ ace-server  (ACE-Step 1.5, GPU)
                                                                  ├─ Ollama      (Qwen 2.5 or your own model — brief & lyrics)
                                                                  ├─ cover model (FLUX.2 klein / Z-Image Turbo)
                                                                  ├─ quality check (Audiobox Aesthetics, CPU, optional)
                                                                  ├─ MIDI image    (Demucs + Basic Pitch, when idle, optional)
                                                                  └─ ffmpeg · audiowmark (mastering, tags, watermark)

The app server only brokers jobs; the bridge does the work and starts each model only while it's needed. Nothing listens on your network — everything binds to 127.0.0.1.

Using your music

• Licences: ACE-Step 1.5 (MIT; its model card explicitly allows commercial use of generated music), FLUX.2 [klein] 4B and Z-Image Turbo (Apache-2.0). See THIRD_PARTY.md.
• Copyright: in many countries purely AI-generated content is not protected by copyright. Your own contributions — lyrics you write, your selection and mastering — can be. The release package records them.
• Distribution: many distribution services accept AI-assisted music if you disclose it; the release package contains ready-to-tick AI credits. Some platforms and music libraries don't accept AI-generated music — check their rules first.
• Self-played tracks (Make your own, or recorded in the loop machine) are your own work: the tags, the release package and Fahrstuhlmusik say “self-played”, and only the cover is declared as AI.
• AI takes over results are AI tracks — the note records which part is yours (your bars or your voice).
• This is not legal advice.

Uninstall

Quit the launcher, then delete ~/ManuwaveStudio (and on macOS, optionally, the Homebrew packages listed in install/macos/install.sh). Nothing else is installed.

Licence

Source-available, not open source — see LICENSE. You may run it for yourself; you may not redistribute or modify it without permission. Music you make is yours.