A convincing creature vocalisation implies a body: lungs of a certain size, a throat of a certain wetness, teeth or a beak or neither. Prompting the anatomy gives specific, repeatable results. Prompting menace gives generic ones, because menace is the audience's conclusion and has no acoustic signature of its own. Describing the mechanism instead of the mood pays off more in this collection than in any other.
Every clip shows the exact prompt that produced it. The render script reads its prompts from the same file as this page. Copy one, change a word and generate your own.
enormous creature roaring, deep chest-driven bellow with a rasping throat and a long tail, open outdoor space
large predator growling a low continuous warning, wet throat rattle with breath behind it, very close mic
small creature shrieking in alarm, piercing high cry breaking into a rasp, echoing cave
insectile creature chittering rapidly, dry clicking mandibles with a thin buzzing vocal underneath, close mic
large creature breathing slowly and wetly, thick mucous rattle on each inhale, very close and intimate
swarm of small creatures approaching, dense overlapping chittering and scrabbling claws on stone, tunnel reverb
creature dying, wet gurgling exhale collapsing into silence, close mic, dark space
huge wet tentacle dragging across stone, thick suction with slapping impacts, cavernous and damp
creature feeding, wet tearing of flesh with bone cracking and heavy breathing between, close and visceral
unknown creature calling from far away, long mournful cry echoing across a vast valley at night
These models respond to physical description. Name the material, the action and the space.
enormous creature roaring, deep chest-driven bellow with a rasping throat and a long tail, open outdoor space
"Chest-driven" implies lung volume, the only reliable cue for size. A small roar pitched down only sounds pitched down.
large predator growling a low continuous warning, wet throat rattle with breath behind it, very close mic
A quiet sustained growl close to the microphone is more threatening than a roar, and "breath behind it" makes it sound alive.
insectile creature chittering rapidly, dry clicking mandibles with a thin buzzing vocal underneath, close mic
"Dry" and "mandibles" push away from the wet-throated mammal the model defaults to on the word creature.
large creature breathing slowly and wetly, thick mucous rattle on each inhale, very close and intimate
Wetness is the most reliable single control for how disgusting a creature is, and it is independent of size and pitch.
"Scary monster" is an instruction the model cannot act on, because no spectral property corresponds to scary. "Wet throat rattle", "chest-driven bellow", "clicking mandibles" and "mucous inhale" are things a body does, and they give specific, repeatable results. Whenever you are about to write an emotion, write the anatomy that would cause it.
Wet creatures have throats, mucus and soft tissue; they gurgle, rattle and slather, and they read as flesh. Dry creatures click, rasp and rustle, and read as chitin, bone or something desiccated. That one axis covers most of the range between a dragon and a swarm. It is independent of size and pitch, so you can move along it without changing your other choices.
A roar close to the microphone reads as a large animal. The same roar heard across a valley with a long echo reads as something enormous, and as something the characters cannot see. Horror gets more out of the second. Put the distance in the prompt ("calling from far away", "echoing across a vast valley") instead of adding reverb, because real distance removes the high end and stretches the timing in a way a plate cannot fake.
The models run on your machine, which changes the cost, the privacy and what you can automate.
Weights are fetched once. After that, generation, the library and the API all work with the network off. Everything lives in one SQLite file and one folder of WAVs under ~/.soundbox, and nothing is uploaded.
Soundbox runs an MCP server in the same process, with tools for generating and listing sounds. Generated files come back as absolute paths, so an agent can copy them into a project.
Every model's terms are shown before you download it, and non-commercial weights need an explicit confirmation. Soundbox is MIT. The licence of the model that made a sound governs what you may do with it.
claude mcp add --transport http soundbox http://127.0.0.1:17494/mcpcurl -X POST http://127.0.0.1:17494/api/generate \
-H 'Content-Type: application/json' \
-d '{"model":"audiocpp-stable-audio-3-sfx","prompt":"8-bit coin pickup blip","duration_s":1}'