Generated locally

Music you generate on your own machine

Tracks rendered on a single desktop GPU by open-weight models, each with the exact prompt behind it to copy, edit and re-render. No account, no API key, no per-track billing, and nothing leaves your machine.

Prompt, lyrics, render, stems

How a track on this page gets made

Two models, split by one capability: Stable Audio 3.0 Small Music has no vocal path, so it makes the instrumentals; ACE-Step takes lyrics, so it makes the sung tracks. Pick a path in the panel and the rest of it follows. Both models are listed on the model page.

Sound: music, from a prompt to an arrangement

ACE-Step 1.5 Turbo: lyrics-conditioned, so the words are their own input.

Prompt

triumphant power pop victory anthem, bright electric guitars, punchy live drums, gang vocal chorus, major key, 140 BPM

Lyrics

[chorus] We took the long way home We took the crown Victory lap, victory lap We are not coming down

A cross-section of the catalogue, from synthwave to orchestral hybrid to sung power pop, all rendered on one desktop GPU.

24 contemporary scenes, from amapiano and melodic drill to corridos tumbados, hyperpop and UK garage. Same two models, twenty-second renders and one GPU as the rest of the page.

Instrumental

20 tracks from Stable Audio 3.0 Small Music: loops and cues built to sit under gameplay, dialogue and film.

Sung with lyrics

16 tracks from ACE-Step, the lyrics-conditioned model: hand it words and it sings them. Every track page shows the exact lyrics behind the take.

Soundbox Records

A fictional label where every artist is a prompt.

Fictional roster

Twenty fictional acts, each with a house style. To a generative model an artist identity is a vocabulary held constant, so every act on this roster is a persona string, and its signature track was rendered from that string. Press play to hear the act. Copy the persona and append your own brief to render in its style.

  • SBX-001

    Chrome Sunset Division

    Synthwave

    wide analog saw leads, gated reverb drums, sidechained bass, neon 1984 palette, 108 BPM

  • SBX-002

    Pixelheart Parade

    Chiptune pop

    bright square wave leads, pulse wave harmony, noise-channel drums, relentlessly major key

  • SBX-003

    The Slow Cavern

    Ambient

    glassy bell arpeggios, long dripping-water reverb tails, sparse hand percussion, minor and unhurried

  • SBX-004

    Foundry Choir

    Industrial

    scrap-steel percussion locked to the grid, distorted bass ostinato, machine-room air, 128 BPM

  • SBX-005

    Battalion of Strings

    Orchestral hybrid

    staccato low string ostinato, brass stabs, taiko on the downbeat, sustained choir pad underneath

  • SBX-006

    Dust & Circuit

    Lo-fi

    dusty vinyl crackle, mellow square wave melody, warm sub bass, laid-back swung drums, 82 BPM

  • SBX-007

    Lanternwood

    Fantasy folk

    plucked lute, wooden flute counter-melody, hand drum and tambourine, warm firelit major key

  • SBX-008

    Low Lamp Trio

    Blues

    weeping guitar bends, brushed drums, warm upright bass, smoky late-set minor key, 68 BPM

  • SBX-009

    Gravity Well Ensemble

    Space ambient

    evolving analog pads, distant reverb-drenched piano, soft granular texture, no percussion

  • SBX-010

    Redline Prophets

    Racing rock

    distorted guitar riffs, four-on-the-floor kit, aggressive bass synth, rising tension, 165 BPM

  • SBX-011

    The Glass Arcade

    Dream pop

    washed-out shimmer guitars, soft airy female vocal, slow chorus pedal sway, reverb over everything

  • SBX-012

    Copper Kettle Club

    Ragtime jazz

    upright piano stride, muted trumpet, brushed snare shuffle, tea-room cheerfulness, 96 BPM

  • SBX-013

    Static Bloom

    Shoegaze

    wall of fuzzed guitars, buried melody, slow crashing drums, everything blooming into feedback

  • SBX-014

    Moth Orchestra

    Chamber

    close-mic string quartet, fluttering pizzicato, hushed dynamics, candlelit minor key

  • SBX-015

    Bitrate Bandits

    Breakcore chiptune

    chopped noise-channel breaks, detuned pulse stabs, frantic tempo shifts, gleeful 8-bit chaos

  • SBX-016

    Velvet Modem

    Synth funk

    slap bass synth, clavinet stabs, tight drum machine groove, talkbox-adjacent lead, 104 BPM

  • SBX-017

    Iron Lullaby

    Doom folk

    down-tuned acoustic drones, funeral-pace percussion, low male hum, heavy and strangely tender

  • SBX-018

    Paper Satellites

    Indie pop

    jangly electric guitars, handclaps, warm synth pads, cheerful duet vocals trading lines, 124 BPM

  • SBX-019

    The Hologram Choir

    Synth choral

    stacked vocoder harmonies, cathedral reverb, slow chord swells, sacred music from no century

  • SBX-020

    Ember Fields

    Post-rock

    clean tremolo guitars building to a distorted crest, cymbal swells, ten-minute patience in twenty seconds

How to prompt for music

Three rules with the largest effect on the output.

Name the instruments and how they are played

“Epic battle music” gives a model nothing to voice. “Staccato low string ostinato, brass ensemble stabs, taiko drum hits on the downbeat” names each part and how it is played, and the render comes back as an arrangement. Naming the parts makes the biggest difference between a first take you keep and one you re-roll.

Set the tempo in BPM

A BPM number moves the output further than “fast”, “driving” and “intense” together, because the model has nothing to interpret. Set it whenever rhythm matters (165 for a racing cue, 68 for a slow blues) and use adjectives for timbre.

Keep lyrics short enough to fit the render

ACE-Step sings the lyrics block verbatim, and a 20-second render has room for a handful of lines. Give it a full verse and it rushes the syllables or silently drops the tail. Use two or four short repeated lines, tagged [verse] or [chorus]. Put the voice description in the prompt: the lyrics field is read only as words to sing.

Why generate music locally

Rendering on your own GPU needs no account, API key, upload or per-track billing.

Offline once the weights are downloaded

Both models run through audio.cpp, a native binary, from quantised GGUF files on your disk. One SQLite file and one folder of WAVs under ~/.soundbox. Nothing is uploaded.

An agent can drive it

Soundbox runs an MCP server. An agent calls soundbox.generate with a prompt and a lyrics block and gets back an absolute path to a WAV. This page was produced through the same HTTP API.

Licences stated before you download

Every track names its model, and every model's terms are on this page. No per-track billing, and Soundbox itself is MIT.

Licensing, in full

Every track names the model that made it. These are both models’ terms, in full.

  • Stable Audio 3.0 Small Music

    audiocpp-stable-audio-3-small-music

    Stability AI Community License: Commercial use is free below USD $1M annual revenue; above that Stability requires an Enterprise licence.

  • ACE-Step 1.5 Turbo

    audiocpp-ace-step

    Apache 2.0: Listed as Apache 2.0 in the catalogue and flagged unverified there. Read the model card before shipping output commercially.

Questions

Can I use this music in a commercial game or video?
It depends on which model rendered the track; every row names it. The instrumentals come from Stable Audio 3.0 Small Music, under the Stability AI Community License: free for commercial use below USD $1M annual revenue, Enterprise licence required above it. The sung tracks come from ACE-Step, listed as Apache 2.0 but flagged unverified in the model catalogue, so read its model card before shipping. Soundbox itself is MIT and imposes no terms on what you generate.
How long can a generated track be?
Stable Audio 3.0 Small Music renders up to 120 seconds in one pass and ACE-Step up to 60. Everything on this page is about 20 seconds, loop or cue length. For anything longer than the limit, generate sections and arrange them.
Which model should I use for music with vocals?
ACE-Step. Stable Audio has no vocal path and cannot sing words; ask it for a chorus and you get an instrumental with a lead line where the voice would be. ACE-Step takes a separate lyrics field and sings it verbatim, and the prompt sets genre and instrumentation.
Does this run in the cloud?
No. Both models run on your own machine through audio.cpp, a native binary, from quantised GGUF weights. There is no Python stack, account, API key or per-track billing. Once the weights are downloaded it works with the network off. The tracks on this page were rendered this way on one desktop GPU.
Why do the durations not match what was requested?
Every track here asked for 20 seconds and came back between 10.7 and 20.0. Diffusion audio models render to a latent length that maps only approximately to seconds, so the requested duration is a target. If you need an exact length, generate longer than you need and trim.
Can an agent generate these instead of me?
Yes. Soundbox exposes an MCP server, so a coding agent can call soundbox.generate with a prompt and a lyrics block and get back an absolute path to a WAV it can drop into a project. Every track on this page was produced through the HTTP API that server is built on.

More game soundtracks, and the rest of the catalogue.

Soundbox Records is a fictional label: every act on the roster is a prompt persona, every cover on this page is generated illustration, and no real artist, release or record is depicted. The audio was rendered locally by the models named above and encoded to MP3 for the web; each file downloads in full. Soundbox on GitHub.