UI Audio Theme

UI Audio Theme Generator

Generate cohesive sets of subtle, minimal UI sound effects using ElevenLabs text-to-sound-effects API. Create "audio themes" — coordinated sets of sounds that share a common aesthetic and map to standard UI interaction constants.

Requires ELEVENLABS_API_KEY set in the environment. See README.md for setup instructions.

Delivery paths

This skill produces UI audio two ways, chosen per event rather than per app:

• ElevenLabs samples (the default pipeline below) — AI-generated .mp3/.wav files for bespoke or branded cues, ambient and voice audio, and any target that needs a shippable file (game/TV/desktop, per design-game-ui).
• Synthesized web micro-interactions via cuelume — deterministic, professionally tuned Web Audio sounds wired in two lines, for the high-frequency, low-stakes clicks (hover, press/release, toggle, success, error) that make a web UI feel responsive. See references/cuelume-web-delivery.md for when to reach for it and how to wire it, and assets/cuelume-event-map.json for the per-event fit.

Before choosing sounds for any event on either path, pick a named moment from references/interaction-taxonomy.md. Its 14-moment vocabulary and restraint rules — a distinct sound per moment, the hover throttle, paired press/release, silent fallback — are production-agnostic and apply whether the clip came from ElevenLabs or cuelume.

Relationship to game and television UI

This skill is the single owner of existing-product audio audits, semantic event-to-sound maps, UI-sound candidate generation, visual audition, non-destructive editing, normalization, accepted asset manifests, and reusable sound-slot or picker improvements. design-game-ui owns the upstream semantic feedback-event map for purpose-built game/TV interfaces: what each event means, exactly when it fires, its required visual equivalent, repetition policy, accessibility, and acceptance criteria.

When invoked from design-game-ui, consume that event map rather than redesigning navigation or inventing raw-key-triggered sounds. Frames owns audio production; the application engineer owns runtime playback lifecycle; the game-UI lead accepts the integrated behavior. voice-clone is separate and should be used only when the requested asset is spoken voice.

cuelume-backed web micro-interactions (references/cuelume-web-delivery.md) are web-only — they play through the browser's Web Audio API and never produce a file, so they are out of scope for the game/TV/desktop file targets this section describes. Those targets stay on the ElevenLabs pipeline, which writes files.

Workflow

Step 0: Audit an Existing Product

When an application already exists, or a new auth/payment/blockchain feature landed after its theme was created, run the existing-app audit before generating or rewiring sounds. Follow references/audit-and-wiring.md, copy assets/event-map-template.json, and validate the completed map:

python scripts/audit_theme.py ./audio-event-map.json \
  --theme ./public/audio/ui/theme.json

Inventory semantic state transitions across routes, overlays, async work, auth, payments, blockchain operations, keyboard, and gamepad input. Map each event to a sound or documented intentional silence. Treat unused theme slots as information, not a reason to invent noisy wiring. Require both source evidence and audible browser evidence. Fix event boundaries before judging the assets.

Step 1: Discover the UI Vibe

Before generating sounds, understand the application's aesthetic:

Application Context:

• What type of app? (wallet, dashboard, social, productivity)
• Primary user emotion? (trust, delight, focus, calm)
• Professional or playful interface?

Audio Direction:

• Preferred tone? (warm/organic, clean/digital, retro/nostalgic, futuristic)
• Reference sounds? (Apple-like clicks, game bleeps, banking chimes)

Step 2: Select a Vibe Preset

Available presets in assets/vibe-presets.json:

Step 3: Generate the Audio Theme

# Using preset vibe
python scripts/generate_theme.py \
  --vibe "crypto-modern" \
  --output-dir "./audio-theme"

# Using custom vibe description
python scripts/generate_theme.py \
  --vibe-custom "warm organic subtle wooden interface sounds" \
  --output-dir "./audio-theme"

# Generate specific categories only
python scripts/generate_theme.py \
  --vibe "crypto-modern" \
  --categories buttons feedback transactions \
  --output-dir "./audio-theme"

# List available presets
python scripts/generate_theme.py --list-vibes

# List all sound names
python scripts/generate_theme.py --list-sounds

Step 4: Regenerate Individual Sounds

python scripts/generate_theme.py \
  --regenerate "button-click-primary,notification-success" \
  --vibe "crypto-modern" \
  --output-dir "./audio-theme"

Step 5: Audition and Refine (Interactive Picker)

Taste can't be automated — after generating a baseline, let the user audition each slot and iterate. Two strategies; pick by runtime:

A. Local picker server (default — works in Claude Code AND Codex):

python scripts/sound_picker.py --vibe pixel-minimal --output-dir ./audio-theme
# → http://127.0.0.1:7777

Serves a local page with, per sound slot: a play button for the current file, an editable generation prompt (pre-filled from the vibe), "Generate next" for unlimited fresh candidates via ElevenLabs (auto-normalized), and compact always-open waveform cards. The selected take uses a persistent visual state, and clicking any card's waveform opens the shared editor. Compact icon controls support:

• accept the candidate into its current event slot;
• assign a good wrong-slot candidate to any other event without regenerating;
• open a visible waveform editor with draggable trim handles, click-to-seek, selection playback, and keyboard-adjustable trim boundaries;
• create a non-destructive revision with plainly labeled fade-in, fade-out, volume, reverb, delay-time, and delay-amount sliders; and
• delete disposable candidates and their derived revisions without deleting the accepted live theme file.

Every edit creates a new revision. Accepting a replacement snapshots the prior accepted file into picker history first, so returning to an earlier version is an ordinary audition-and-accept operation. The picker records prompt, source slot, edit recipe, assignment provenance, and history in theme.json. Run it in the background, hand the user the URL, and continue only after explicit per-slot acceptance. No CSP constraints apply because it binds to localhost.

The picker is the default review interface, not an optional polish step. Give the user the local URL and wait for explicit per-slot selections. Do not treat the first generated baseline as accepted, and do not substitute filenames or a written description for listening. After acceptance, return the recorded winning prompts and event-to-file manifest to Frames or the implementation owner.

B. Artifact audition board (Claude Code only, when a hosted page is preferred): hosted Artifacts enforce a strict CSP — no external requests, so live generation inside the artifact is impossible. Instead: pre-generate 2-4 candidates per slot with --regenerate, embed each mp3 as a base64 data: URI in an HTML page with play buttons and the prompt shown per candidate, publish via the Artifact tool, and let the user reply with picks and prompt tweaks for the next batch round. Iterative rounds instead of infinite Next; keep candidate counts small — base64 audio inflates page size ~33%.

Loudness Normalization

ElevenLabs SFX loudness tracks the prompt wording — "quiet"/"subtle" vibes can come back peaking at -31 dBFS, inaudible once an app applies its own 20-40% UI volume. The generator therefore peak-normalizes every file to -3 dBFS by default (requires ffmpeg; pass --no-normalize to keep raw output). Keep the vibe words for timbre; let the app's volume setting control loudness.

Integration Gotchas

• Autoplay policy: browsers block ALL audio until the user's first click/tap/keypress on the page. Hover-triggered sounds only work after that first gesture — this is a hard browser rule, not an implementation choice.
• Hover cues: play them 40-60% quieter than clicks and throttle (~90ms) so cursor sweeps across menus don't machine-gun.
• Hover semantics: use nav-item-hover for primary navigation/menu browse, item-hover for other interactive items, and tooltip-show only when a real tooltip becomes visible. Do not make tooltip audio the global hover sound.
• Transactions: place tx-sent after a wallet/network returns a broadcast transaction id, tx-pending on the transition into a meaningful waiting state, tx-confirmed on the first confirmed/fulfilled transition, and tx-received only in a receiver-facing UI. Never fire them on the button click that merely starts the operation.
• Editing: keep originals immutable, render revisions through bounded numeric ffmpeg filters without shell=True, leave output headroom, and retain a previous accepted snapshot before replacement. Never normalize an edited file in a way that cancels intentional gain changes.
• Verify audibly, not just mechanically: audio.play() resolving proves playback, not perceptibility. Check peaks with ffmpeg -i file.mp3 -af volumedetect -f null - and listen at the app's actual volume multiplier before shipping.

Sound Categories

Each constant below maps to a named moment in references/interaction-taxonomy.md, which defines its intended character and keeps press vs release and toggle vs tick genuinely distinct. Note: press/release as a pointerdown/pointerup pair are folded into the button-click-* constants today; the taxonomy names them explicitly so a future generator or a cuelume-backed wiring can split them.

Buttons

Navigation

Feedback

States

Modals

Transactions (Wallet-specific)

Output Structure

audio-theme/
├── theme.json              # Theme manifest
├── constants.ts            # TypeScript constants
├── .picker/                # Immutable candidates, revisions, and history
├── buttons/
│   ├── button-click-primary.mp3
│   ├── button-click-secondary.mp3
│   └── button-click-destructive.mp3
├── navigation/
├── feedback/
├── states/
├── modals/
└── transactions/

Script Options

--vibe NAME           Preset vibe name
--vibe-custom DESC    Custom vibe description
--output-dir PATH     Output directory (default: ./audio-theme)
--format FORMAT       mp3 or wav (default: mp3)
--categories CATS     Specific categories to generate
--regenerate SOUNDS   Comma-separated sounds to regenerate
--prompt-influence N  0-1, higher = stricter prompt adherence (default: 0.5)
--list-vibes          Show available presets
--list-sounds         Show all sound names

Resources

• scripts/generate_theme.py — CLI tool for generating themes
• scripts/sound_picker.py — Local audition, assignment, editing, and history UI
• scripts/audit_theme.py — Validate an existing app's semantic event map against its theme
• references/audit-and-wiring.md — Existing-product inventory, wiring, and evidence workflow
• references/sound-design-guide.md — Detailed sound design best practices
• references/interaction-taxonomy.md — The 14-moment vocabulary + restraint rules (production-agnostic)
• references/cuelume-web-delivery.md — Synthesized web micro-interaction delivery via cuelume
• assets/vibe-presets.json — Predefined vibe configurations
• assets/cuelume-event-map.json — Our semantic events mapped onto cuelume's 14 sounds
• assets/theme-template.json — Example output manifest
• assets/event-map-template.json — Auditable interaction-to-sound map template
• README.md — Prerequisites, design philosophy, integration examples (React hook, Howler.js), accessibility guidance