Sound Design is the x-rush.ai workbench for the audio nobody sings and nobody speaks: door slams, rain on glass, UI clicks, room tone, the whoosh under a title card. It shares the Audio module with Music Gen and splits the territory cleanly. Songs and instrumental tracks come from Music Gen; narration and cloned voices belong to Voice Studio. Everything else a scene asks you to hear is this feature's job.
The operation fits in one sentence. Describe a sound, optionally attach one reference clip, confirm, and a finished effect or ambience bed lands in your Project as a downloadable MP3. Every confirmed generation is one ordinary audio unit at 2 credits, and every generation returns exactly one sound. The price sits printed on the generate button, and pressing it submits the task; no preview step and no second dialog stand between you and the run.
Two Ways In
The feature opens with two doors. The curated library hands you a usable sound without generating anything, free of charge. Custom generation builds the sound you described, at the fixed price. Both paths end in the same place: a private Sound Design result inside the Project you were working in.
The Curated Sound Library
The library is a platform-run catalog fixed at eight categories: People & Actions, Nature & Weather, Animals, City & Traffic, Indoor & Everyday, Machines/Tech/UI, Impacts & Transitions, and Ambience. Each entry lists its name, duration, category, a few tags, whether it loops, and its usage terms before you commit to anything. Search matches platform-written names and hand-set tags; there is no semantic search guessing at "something rainy and sad". Filters cover sound type, category, duration range, and loopability, and the default ordering is human curation rather than an engagement algorithm.
Auditioning is free and one sound plays at a time across the page; starting a second stops the first and resets it. Copying an entry into your Project is also free, creates a standalone private result with no workspace behind it, and consumes neither an AI generation nor an agent assist. Rights are settled before listing, not after: every entry carries one reviewed label, usable in commercial projects with no attribution required, and sounds whose licenses demand per-file credit never make the catalog at all.
When the entry is close but the scene needs its own version, Generate Similar turns it into a reference for custom generation at the normal 2-credit price. The click snapshots the sound as a private reference, opens a new workspace, and leaves the library entry untouched.
Custom Generation
A workspace holds one sound intent: the sound type, the written description, duration, loop setting, and the full history of generations run against it. Single effects run from 0.1 to 30 seconds. Ambience runs from 1 to 30 seconds and is the only type with a seamless-loop option; effects have no loop switch. The system proposes a default duration from your description, and the number stays editable until you confirm.
Each generation takes the text plus at most one optional reference clip. Four sources feed it: a direct upload in MP3, WAV, M4A, or OGG (up to 100 MB and five minutes of decoded audio), a file from your account media library, an entry from the curated library, or an existing result from any of your Projects. The reference steers the target's material, rhythm, and sense of space; it does not clone the source, and no result is promised to match it sample for sample. Anything longer than 30 seconds must be trimmed by you, down to one continuous clip, with a trimmer that offers a start point, an end point, playback, and a confirm button. The system never cuts audio on your behalf. Uploads serve the current task by default, and an opt-in checkbox, off until you tick it, adds the file to your media library. The server validates what the file actually contains rather than trusting its extension, so a renamed file does not sneak through. Video files are not inputs here, and the feature does not extract audio tracks from them.
While a reference uploads, validates, or trims, the paid submit stays blocked; if the preprocessing fails, your description and settings survive and you can retry or swap the clip. Nothing about a failed upload ever creates a task or reserves credits.
Working with the Sound Design Agent
The agent does three things, all on request. Optimize Prompt rewrites your description into a fuller generation note, adding material, space, distance, and intensity for an effect, or scene layers and a loop decision for ambience. It analyzes a sound event you describe. It also deals out random combinations of material, space, and sound events for the moments you have no direction at all. None of this fires on its own, and none of it can start a paid generation.
Clicking Optimize Prompt and receiving the editable note counts once on the account-wide assistance pool, where ten successful assists bill 0.5 credits in one charge. The completion work that happens inside a generation itself, filling whatever gaps the task needs and proposing a workspace name with up to three tags on your first confirmed run, is part of the 2-credit price and never touches the pool. If that naming suggestion fails, the workspace simply becomes Untitled sound plus a timestamp, and the generation goes ahead anyway.
What a Generation Costs
Two credits, every time, for effects and ambience alike, with or without a reference clip. Credits are reserved when the backend accepts the task and consumed only when the sound and its MP3 are both durably saved. System retries inside a task never charge twice, and a task that cannot fully deliver releases the reservation in full. Regenerating is a fresh decision every time, 2 credits again, and each success appends a new result to the workspace history; nothing is ever overwritten. One confirmed task returns one sound, so comparing takes is on you, at 2 credits per take.
One workspace runs one task at a time, and while it runs you can draft the next version of the description, though the running task keeps the inputs it was accepted with. How many tasks can run at once depends on your plan: one on Free, two on Lite, three on Pro. Accepted tasks beyond that wait in the queue with their credits still reserved. Close the tab, refresh, or switch devices and accepted tasks keep running server-side; the workspace restores its state when you come back. Switching Projects saves your drafts first, asks before cancelling an in-flight upload, and leaves accepted tasks with the Project that owns them.
What You Get Back
Delivery is all-or-nothing. A result exists only when the audio generated, the MP3 conversion passed verification, and the whole record saved reliably; until then you see task status, never a half-rendered sound. There is exactly one file per result, a single MP3 picked for direct use. WAV, sample rates, bitrates, and channel layouts are not exposed, and no separate export step exists to pay for.
Playback follows one rule per page: start a sound and anything else playing stops and resets to its beginning. Only ambience generated with the seamless-loop requirement gets a loop control; everything else plays once through and ends. Downloading hands you that same MP3 in one click. A result that delivers completely counts as delivered even when it is not the sound you heard in your head. Delivery checks ask whether the sound arrived complete; how it sounds is your judgment, and the remedy is another 2-credit attempt with edited input.
Results inherit the workspace name and tags at birth and can be renamed independently afterward. Finding them again goes by name, tag, or sound type, and there are no draft-versus-final stages, folders, or favorites to manage. Continue Adjusting reopens the workspace that produced a sound, right at that generation's note. Generate Similar always opens a new workspace with the sound as reference. Taking a sound into another feature happens from that feature's side: its selector picks from Sound Design results, defaulting to your current Project, and Sound Design itself has no send-to button. The mobile layout keeps playback and the MP3 download one tap away and folds adjust, similar, publish, edit, and delete into an action menu; the rules underneath are identical on both.
Deletion splits along the same lines and never comes back. A workspace and a finished result are deleted separately, immediately, and without a recycle bin; neither is reversible and neither is refunded. Deleting a workspace clears its drafts and history but leaves every delivered result in place. Deleting a result destroys the sound and its MP3, retires its Explore entry and Remix entry, and leaves finished Remixes that other people made untouched. A workspace with a task still running refuses to delete until the task ends.
Publishing to Explore
Publishing takes one result and freezes it into a public snapshot: the audio, title, description, tags, sound type, duration, loop status, your public attribution as of publish time, and the final prompt only if you say so. Drafts, reference clips, and the workspace history stay private regardless. Nothing publishes automatically; a finished generation is not a publication.
Three independent switches, all off by default: Allow Download, Allow Remix, and Public Final Prompt. Downloads, once enabled, serve a watermarked public copy (your author name plus x-rush.ai) straight from the detail page, and switching them off later stops new downloads without recalling files already taken. The watermark marks delivery permissions; it is not DRM and does not disable a browser cache or a system recorder. Remix gives a visitor a private reference snapshot in their own Project and a brand-new workspace, and they inherit your prompt only when you published it. Playback stays inside the Explore detail page: no standalone work page, no raw file link, and no share-link actions. AI transparency labeling follows the target market and channel rather than one global rule.
What Sound Design Does Not Do
The boundaries are specific, and each one has a reason:
- No songs or background music. Complete tracks and instrumentals are Music Gen's territory, including the BGM shortcut covered on the BGM page.
- No voices. Narration, speech, and cloned voices live in Voice Studio, where voice cloning carries its own consent rules and its own honest status page.
- No automatic Foley or multitrack. One sound per task. Spotting effects against a video timeline and generating themed sound packs in bulk are outside the feature.
- No variant picker. One confirmed generation returns one sound. The workspace history keeps every take you paid for, but choosing between them is your work.
- No template system. The curated library is platform-run. There are no user sound libraries, no submissions, and no shareable generation templates.
- No editing. No mixing, denoising, or splicing. The reference trimmer, with its start point, end point, preview, and confirm, is the entire extent of audio manipulation here.
That last line is the operating philosophy in miniature. The feature generates, delivers, and gets out of the way; it does not score your sound, grade your taste, or hold your files hostage behind extra export fees. Where it stops, the neighbors start: Music Gen for anything with a melody, Voice Studio for anything with a voice. And if the sound you need already exists in the catalog, it costs nothing at all.