Horror sound effects
Models ignore mood labels like "scary". Describe the physical event: metal under strain, a room too large for its sound.
Listen: 11 samples from the horror collection
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.
Whispered dread chant
via audiocpp-ace-stepdark ambient horror theme, whispered ghostly vocals, dissonant strings, unsettling drone, slow tempo
Music box
via audiocpp-stable-audio-3-sfxantique music box playing a slow detuned lullaby, warped tape wow and flutter, dusty attic reverb
Bone crack
via audiocpp-stable-audio-3-sfxbone snapping and cartilage tearing, wet crunching with a sharp final crack, close and dry
Dragging sound
via audiocpp-stable-audio-3-sfxsomething heavy being dragged slowly across a wooden floor, scraping with intermittent stops, empty house
Distorted voice drone
via audiocpp-stable-audio-3-sfxunsettling drone made of stretched distorted human voices layered and detuned, slowly swelling
Nails on metal
via audiocpp-stable-audio-3-sfxfingernails scraping slowly down a rusted metal sheet, piercing high scrape with a rattling tail
Panic heartbeat
via audiocpp-stable-audio-3-sfxhuman heartbeat accelerating from slow to frantic, deep thumping pulse with chest resonance, close mic
Footsteps above
via audiocpp-stable-audio-3-sfxslow deliberate footsteps on floorboards heard from the room below, muffled creaking, tense spacing
Swarm approaching
via audiocpp-stable-audio-3-sfxswarm of insects approaching, dense buzzing rising in volume and density with erratic close flybys
Dissonant organ
via audiocpp-stable-audio-3-sfxpipe organ holding a dissonant minor cluster chord, slowly swelling in a vast cathedral
Whisper in static
via audiocpp-stable-audio-3-sfxradio static with faint unintelligible whispering emerging and receding, tuning drift, close mic
Prompts that work
These models respond to physical description. Name the material, the action and the space.
Metal groan
metal under extreme strain, slow deep groaning creak, low frequency, no percussive attack, 3 seconds, roomy and unsettling
"Metal under extreme strain" describes a physical event. The model can render strain; it can do nothing with "scary".
Room too large
sustained low drone in a room too large for its own sound, faint resonant hum, no melody, 4 seconds, distant and cold
The model renders "a room too large for its own sound" as thin, distant resonance, which sounds uneasy.
Jump stinger
sharp horror stinger, sudden dissonant metallic hit, high-frequency shriek, instant attack, hard cutoff, 500 ms
"Instant attack" and "hard cutoff" make a stinger hit as a shock. A slow fade-in ruins a jump scare.
Scratching behind a wall
faint scratching sound behind a wall, irregular rhythm, dry scraping texture, low volume, distant and muffled, 2 seconds
"Irregular rhythm" keeps this from sounding mechanical. A perfectly even scratch sounds like a machine.
Close breathing
slow uneven breathing close to the microphone, wet and ragged texture, irregular rhythm, no words, 3 seconds, intimate and dry
"Close to the microphone" plus "dry" removes the room around the sound, so the breathing seems to be right next to the listener.
Making horror sounds that fit
Describe the physical event
Models ignore mood labels like "scary", because training data is tagged with things like metal, strain, a room or breathing. Translate the feeling into a physical event: dread becomes "metal under strain" or "a room too large for its sound", tension becomes "irregular rhythm", a jump becomes "instant attack, hard cutoff". The mood follows from the physics.
Set the space
Horror sound design depends on room and distance more than most categories. The same groan is unremarkable close and dry, and unsettling when it is "distant and muffled" behind a wall or in "a room too large for its own sound". Decide where the listener is relative to the source before you write the prompt. Proximity and space often matter more than the source itself.
Make living things irregular
A perfectly even scratch, drip or breath pattern sounds like a machine or a loop, and is less frightening than something unpredictable. Write "irregular rhythm" or "uneven" for anything meant to suggest a presence. That instruction separates background ambience from something that seems alive.
Use a whispered line sparingly
Every other sample on this page is texture: a groan, a drone, a scratch, with no voice. ACE-Step can render a whispered line ("[verse]\nThey are near\nDo not look"), and a barely intelligible chant suggests a presence that knows the listener is there. Use it once per scene at most, and keep it out of background beds.
Why generate locally
The models run on your machine, which changes the cost, the privacy and what you can automate.
Offline after download
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.
Your agent can use it
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.
Licences stated up front
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}'Frequently asked questions
- Can I use these sounds in a commercial video or game?
- It depends on the model that made them. Soundbox shows each model's licence before you download its weights. The Stable Audio 3 models use the Stability AI Community License: commercial use is free below USD $1M annual revenue, and above that Stability requires an enterprise licence. ACE-Step 1.5 is MIT. MiniMax's models allow commercial use under a revenue cap and require an attribution notice. Some voice and stem-separation models are non-commercial only. Soundbox itself is MIT, but the audio is governed by the licence of the model that made it. Read that licence before you ship; this summary is not legal advice.
- Does this run offline?
- Yes, once the weights are downloaded. Soundbox runs the models locally and stores everything in one SQLite file and one folder of WAVs under ~/.soundbox. There is no account, API key or upload step. Only the initial model download needs the network.
- What hardware do I need?
- An NVIDIA GPU with 8 GB or more. CPU-only is supported, but each generation then takes minutes. That is fine for a few short clips and slow for a large batch.
- Can my coding agent generate these for me?
- Yes. While Soundbox is running it serves MCP at http://127.0.0.1:17494/mcp, with tools for generating and listing sounds. An agent can generate a sound as part of the task it is already doing and gets back an absolute file path to use in your project.
- How is this different from downloading a sample pack?
- A horror pack gives you someone else's stingers and drones, made to fit any scene. Generating lets you tune the physical event (which metal, how large the room, how ragged the breathing) to your scene, and regenerate the stinger until the timing works. It costs time and a GPU.