Sound Effect Generator
Type a description, get a deterministic synthesized WAV — 100 cues with editable JSON specs
100% In-Browser · Private & Instant · No Signup
Your file is decoded, processed, and re-encoded in this browser tab. Nothing is uploaded to an OfflineTTS processing API.
Interactive Workspace
Sound Effect Generator
Advertisement
Adsterra privacyDescribe a sound in plain words and this page synthesizes it: no sample library, no upload, no account. Type “coin pickup”, “rain and thunder”, or “door slam in the rain, then a low rumble”, and the planner picks cues from a fixed library of 100 effects, writes a JSON spec, and renders the audio in your browser.
What you get back: one stereo 48 kHz WAV at 16-bit, leveled toward a consistent RMS with a peak ceiling so the file has headroom instead of being loudness-maximized. Every cue can be played from the demo library, downloaded on its own, or loaded into the editor to change its duration, pitch, brightness, gain, or noise seed.
How the sound is made: the renderer is plain arithmetic. Oscillators (sine, triangle, saw, square) with pitch glides and envelopes handle the tonal parts; white noise through a state-variable filter handles wind, rain, explosions, and whooshes; short filtered bursts become clicks, footsteps, and crackle. Nothing is fetched and no recording is involved, so a “rain” cue is a synthesis recipe rather than a field recording.
Why the JSON matters: a sound is a spec, not a mystery blob. The spec lists the cue id and its parameters, and the same spec always renders the same samples — the same file, byte for byte. That makes an effect reproducible in a project, comparable between takes, and easy to hand to a collaborator as a tiny JSON object instead of a 400 KB WAV.
Where this is useful: game jams and prototypes that need placeholder SFX with clean licensing; short-form video and podcast edits that need a transition, whoosh, or notification; UI design where a click should be short and consistent; and teaching, where changing one parameter and hearing the consequence is more instructive than browsing a sample pack.
Scenario coverage: the library is grouped by the job the sound does — interface feedback and transactions, alarms and doorbells, game rewards and fail states, weather and wildlife beds, household foley such as zippers, keys, matches and pouring water, place scenes such as an airport terminal, train station, cafe, haunted house, rainy window, night pond and beach, plus cinematic swells and brand intros. Each scene is one cue with its own internal layers, so “haunted house” does not need three planner steps.
Where it is not: this is not a recording, and it will not pass for one. Explosions and door slams sound synthesized, not captured. There are no voices, no music, no stereo placement, no convolution reverb, and no per-file loudness compliance. The caption planner is keyword-based rather than a language model, so unusual descriptions fall back to a neutral pop instead of inventing a new sound. Files never leave the browser tab; the generated WAV is assembled locally and is not stored.
How It Works
Type a description
Use library words such as coin, rain, alarm, whoosh, or explosion. Join two or three beats with “then”.
Generate the plan
The keyword planner builds a JSON spec for each fragment and shows which cue it chose.
Listen and tweak
Play the render, then edit the spec fields and press Render spec to hear the change.
Export the WAV
Download the current render, or grab a single cue from the demo library as its own file.
Features
Caption to Cue
A rule-based planner maps words like coin, rain, laser, or whoosh to a cue and applies loud, low, short, and long modifiers.
Editable JSON Spec
Change the cue, duration, pitch, brightness, gain, room, or seed and re-render from the same engine — no model in the loop.
100 Demo Cues
Eight groups: UI feedback, alerts, games, nature, foley, sci-fi, scenes, and cinematic hits, each playable and downloadable.
Deterministic WAV
The same spec produces identical samples every time, exported as mono 44.1 kHz 16-bit WAV.
How This Caption Planner Maps Words to Cues
The planner is deliberately small and inspectable. It splits a caption on “then”, commas, semicolons, and “and”, keeps up to three fragments, and scores each fragment against the cue library's keyword list. Longer and more specific keywords win: “door slam” beats “slam”, and “school bell” beats “bell”. Modifier words then adjust the spec — “loud” raises output gain, “low” transposes the cue down and darkens its filters, “short” cuts the duration — and an explicit duration such as “3 seconds” overrides the default.
That design has known edges. Keywords match whole words, so synonyms outside the list do not resolve, and a fragment containing two plausible cues picks the higher-scoring one rather than layering them. “Heavy rain on a tin roof” resolves to rain but cannot model the roof; “a robot falling down stairs” has no matching cue and falls back to a neutral pop. The explanation line under the prompt names every cue the planner chose, which is the fastest way to see why a take sounds the way it does.
Deterministic Synthesis, WAV Export, and Credits
Each cue is a small recipe over a stereo sample buffer: band-limited oscillators with pitch glides, noise through resonant biquad filters, short percussive bursts, panning per layer, and exponential or sustained envelopes, followed by a Schroeder room send. The mix is then voiced through a gentle low-pass where a cue needs it, leveled toward a category RMS target under a peak ceiling, given the requested dB gain, and faded at both ends to prevent clicks at the file boundary. The noise seed is part of the spec, so “rain” with seed 1 is always the same rain — change the seed and the droplets move, while the character of the cue stays the same.
Demo cues are pre-rendered from the same engine and published as content-addressed WAV objects on OfflineTTS's R2 asset domain, so the library streams instead of re-synthesizing on every listen; if an object cannot load, the page falls back to local synthesis. The page exchanges ideas with ORB (github.com/stdevMac/orb-audio), an Apache-2.0 caption-to-audio engine whose JSON-spec presentation prompted this workflow. This implementation shares no ORB code, model weights, planner, or audio data; it is a separate procedural synthesizer, and the OfflineTTS application around it remains proprietary. ORB's name and repository belong to its author, and its licence governs its own files, not this page.
Two practical limits follow from the design. First, the stereo field comes from per-layer panning and a small room, not from a full spatial mix; pan or widen it further in an editor if the scene needs it. Second, RMS leveling is not certified loudness normalization. If a destination enforces a LUFS target, measure the exported WAV in a loudness meter and trim it there rather than assuming this render already complies.
Related Tools
Add Reverb to Audio
Place a dry recording in a room, hall, plate, or cathedral
Audio Equalizer
Ten bands from 32 Hz to 16 kHz, with presets and a spectrum view
Audio Noise Remover
Learn the noise floor of a recording, then gate it out band by band
Tone Generator
Test tones, tuning references, and frequency sweeps generated locally
Sound Effect Generator — FAQ
Is this sound effect generator free?
Yes. Playback, spec editing, and WAV downloads are free and need no account. Synthesis runs in your browser, so there is no per-file upload or server quota.
Are these sounds real recordings?
No. Every cue is synthesized from oscillators, noise, filters, and envelopes. The result is intentionally clean and repeatable, but it is not a field recording and will not match a sample library for realism.
Can I use the generated WAVs commercially?
The tool does not add watermarks or usage limits to the audio it renders. Whether a given use is permitted still depends on the platform or client you deliver it to, and on your own rights review — this page cannot grant rights it does not own. Do not present a synthesized effect as a recording of a real event.
Why did my caption produce the wrong sound?
The planner is keyword-based. It scores library keywords in each fragment and picks the best match, then applies modifiers such as loud, quiet, low, high, short, or long. A description with no library keyword falls back to a neutral pop. Use the demo library labels as a vocabulary, or load a cue directly and edit its spec.
What audio format and limits apply?
Output is stereo 48 kHz 16-bit PCM WAV. A render is capped at 20 seconds, and a caption plan holds at most three fragments. Playback and rendering use device CPU; the engine is lightweight but not instant on very old hardware.
Does this use ORB, a neural model, or a sample pack?
No. The engine is an independent procedural synthesizer. The page format — caption, JSON spec, rendered audio — is inspired by ORB by stdevMac, which is Apache-2.0 licensed; no ORB code, model, or data is included here. See the credit line under the workspace.
Which scenarios does the demo library cover?
Eight groups: interface feedback, alarms and doorbells, game rewards and fail states, nature beds (rain, thunder, wind, forest, stream, waterfall, summer night, swamp, beach, snow), household foley (footsteps, doors, keys, zippers, matches, pouring water, typing, camera, glass, applause, phone), sci-fi, place scenes (airport terminal, train station, cafe, haunted house, horror scare, rainy window, night pond, sword fight), and cinematic hits. Place scenes are single cues with layered internals, so one caption returns a complete ambience.
Is my text or audio uploaded?
No. Caption parsing, synthesis, and WAV encoding all happen in this browser tab. The page does not send the caption or the rendered audio to a processing API.
Technical Standards & Upstream References
More Local Audio Tools
Every audio utility here runs in the browser tab. Browse the full set of editing, conversion, and measurement tools.