Sound generation without a key

Three open paths, each right for a different job. Pick by what you need the sound to be, not by what is installed.

Open-weight models

• Stable Audio 3 Small-SFX (stabilityai/stable-audio-3-small-sfx, ~460 M parameters) — Stability's dedicated sound-effects model, sized for CPU inference, up to two minutes per clip. Stability AI Community License: free for research and commercial use under US $1 M annual revenue; above that, an enterprise license. The right default in 2026.
• Stable Audio Open Small (stabilityai/stable-audio-open-small, ~340 M) — the earlier 11-second model, Arm-optimized; generates ten seconds of audio in about seven seconds on a phone-class CPU. Same license.
• Meta AudioGen (AudioCraft) — code is MIT, but the weights are CC BY-NC 4.0. Fine for a prototype, not for a shipped product.

from stable_audio_tools import get_pretrained_model
from stable_audio_tools.inference.generation import generate_diffusion_cond
model, cfg = get_pretrained_model("stabilityai/stable-audio-3-small-sfx")
audio = generate_diffusion_cond(model, steps=8, device="cpu",
    conditioning=[{"prompt": "single very short tick, fingernail on thin plastic, dry, no tail",
                   "seconds_total": 1}])

Prompt exactly as in sound-generation-elevenlabs — material first, "dry, no tail" always — then post-process per sound-spec.

Procedural synthesis

For ticks, taps, pops, and two-note tones, a synthesizer beats any model: the sound is deterministic, weighs nothing, and its pitch and length are parameters you can tie to the UI (a taller card, a lower thud). scripts/sound-family.mjs --provider synth ships six archetypes — tick, tap, chime, thud, pop, whoosh — and writes WAVs from the same manifest the ElevenLabs path uses, so a team without a key gets a coherent family, not silence.

In the browser, the same idea in eight lines is often enough:

// A tick: 60 ms of high sine with an exponential decay. No file, no fetch.
function tick(ctx: AudioContext, freq = 2400, ms = 60) {
  const o = ctx.createOscillator(), g = ctx.createGain(), t = ctx.currentTime;
  o.frequency.value = freq;
  g.gain.setValueAtTime(0.5, t);
  g.gain.exponentialRampToValueAtTime(0.001, t + ms / 1000);
  o.connect(g).connect(ctx.destination); o.start(t); o.stop(t + ms / 1000);
}

ZzFX (MIT, under 1 KB) is the mature version of that idea — twenty parameters covering waveform, envelope, pitch sweep, modulation, and filter — with a visual designer that exports one-liners. Its idiom is 8-bit; tame the square waves and shorten everything for product UI. jsfxr is the browser port of the classic sfxr and lives in the same register.

CC0 libraries

• Kenney — Interface Sounds (100 files) and UI Audio (50 files), CC0, the most-used free UI sets in existence. Which is also the problem: they are recognizable. Use them as a starting family and re-pitch or layer, don't ship them raw.
• soundcn — 700+ short sounds packaged for the shadcn CLI: npx shadcn add @soundcn/click-soft drops a base64-inlined TypeScript module and a useSound hook into your repo. Mostly Kenney-derived CC0; skip its World of Warcraft pack, which is non-commercial.
• Freesound — APIv2 with filter=license:"Creative Commons 0"; token auth for search, OAuth2 for download. The largest pool and the most uneven; budget time to audition.

Whatever the source, the sound-palette rule still applies: one material. A CC0 grab-bag is the audio version of the Lucide-icon tell in ai-default-tells.

When to apply

Any time an installer wants sound and has no ElevenLabs key — or has one and wants deterministic, parameter-driven sounds. sound-family.mjs chooses automatically: key present → ElevenLabs, absent → synth.

Gotcha

"Open source" and "commercial use" are not the same claim. AudioGen weights are non-commercial; Stability's license flips at US $1 M revenue; soundcn mixes CC0 with a proprietary pack. Record the license next to every file in the manifest, and re-check it before a launch.

Sources

• Stability AI — Stable Audio 3 family (May 2026), Stable Audio Open Small model card, Community License terms.
• Arm, Running Stable Audio Open Small with KleidiAI — on-device timing.
• Meta, AudioCraft repository — MIT code, CC BY-NC 4.0 weights.
• KilledByAPixel, ZzFX — parameters, size, license; chr15m, jsfxr.
• Kenney — Interface Sounds and UI Audio packs (CC0); kapishdima, soundcn; Freesound APIv2 docs.
• Related: sound-generation-elevenlabs, sound-playback-web, dependency-discipline.