Vocal effort makes a character sound like a person. It is also the layer most likely to break immersion, because listeners know human voices well and notice a wrong one immediately. Treat a scream as a shape with an intake, a release and a collapse. Prompting the shape gives more usable results than prompting the emotion.
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.
woman screaming in terror, sharp full-throated cry breaking at the top, medium reverberant interior
man shouting in sudden fear, short sharp cry cutting off abruptly, close and dry
short pained grunt from a body taking an impact, air forced out with a strained edge, close mic, dry
person straining to lift something heavy, low sustained effort grunt with a shaky release, close mic
sudden sharp intake of breath, startled gasp then held, very close mic, quiet room
person breathing fast and shallow in panic, ragged inhales with a tight throat, very close mic
warrior roaring a battle cry, full chested and sustained, open outdoor space with a distant echo
someone shouting for help from far away, thin and reverberant, echoing across a large empty space
crowd of people screaming and panicking together, overlapping voices scattering, street reverb
long falling scream receding with a doppler drop as it disappears into the distance, open air
These models respond to physical description. Name the material, the action and the space.
woman screaming in terror, sharp full-throated cry breaking at the top, medium reverberant interior
"Breaking at the top" separates a real scream from a sung note. Voices distort at maximum effort, and listeners expect it.
person straining to lift something heavy, low sustained effort grunt with a shaky release, close mic
Effort is sustained and low; pain is sudden and sharp. Naming which one you want stops the model blending them into something ambiguous.
person breathing fast and shallow in panic, ragged inhales with a tight throat, very close mic
"Tight throat" gives the constriction that sounds like fear. Without it, panicked breathing sounds like exercise.
someone shouting for help from far away, thin and reverberant, echoing across a large empty space
Distance strips the body out of a voice and leaves the reflections. Asking for it directly works better than adding reverb to a close recording.
"Scared scream" gives the model an abstraction; "sharp full-throated cry breaking at the top" gives it a physical event. These models respond better to descriptions of what the body is doing (air forced out, throat tightening, voice cracking, breath held) than to the feeling behind it. The same rule runs through all of Soundbox's prompting advice, and it makes the biggest difference with vocals.
The same voice screaming in a tiled bathroom, a forest and a cathedral gives three different scenes. The reflections arrive after the voice and tell the audience how far away the person is and how big the space is, which dialogue cannot. Always place the vocal somewhere. A dry scream sounds like a sound effect however good the performance.
Games and film use far more grunts, gasps and breaths than screams, and those small sounds keep a character present. They are cheap to generate and easy to vary: three renders of the same effort grunt give enough variation to avoid the repetition that makes game audio grating. Build that library first. You will use it constantly, and the scream rarely.
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}'