Complete ElevenLabs keyboard shortcuts and commands reference — 9 shortcuts across 2 categories. Quick reference cheat sheet for Windows & Mac.
| Shortcut | Action |
|---|---|
| Ctrl + Enter | Generate audio |
| Space | Play / pause |
| Ctrl + Z | Undo |
| Ctrl + S | Save project |
| Shortcut | Action |
|---|---|
| <break time=1s/> | Add pause |
| <emphasis level=strong> | Emphasis |
| <prosody rate=slow> | Change speed |
| <prosody pitch=+2st> | Change pitch |
| <say-as interpret-as=spell-out> | Spell out |
ElevenLabs' generation loop itself is tiny — Ctrl + Enter generates audio, Space previews it, Ctrl + Z undoes, Ctrl + S saves the project. The real decisions are upstream, in where the voice comes from. Instant Voice Cloning builds a usable clone from a small audio sample in moments; Professional Voice Cloning trains on substantially more audio for the high-fidelity result you would put in production. Voice Design skips samples entirely — you describe a voice in words and generate candidates — and the Voice Library offers community-shared voices when you need something serviceable today.
Delivery is steered two ways. Inline tags: <break time=1s/> forces a pause, <prosody rate=slow> and <prosody pitch=+2st> bend speed and pitch, <say-as interpret-as=spell-out> spells acronyms. And generation settings: stability trades consistency against expressiveness, similarity_boost pulls output closer to the source voice, style exaggerates the speaker's manner, and use_speaker_boost sharpens speaker identity.
model: eleven_turbo_v2 optimises for latency — the pick for interactive and streaming uses — while model: eleven_multilingual_v2 favours quality and broad language coverage for produced content. And before reaching for any tag, try punctuation: commas, periods, and ellipses are how the models pace themselves, so well-punctuated text often needs no SSML at all. Writing in ALL CAPS backfires — it reads as spelling or shouting rather than emphasis.
It controls how much the voice is allowed to vary. Low stability produces expressive, emotional reads that can wander between takes; high stability produces consistent, even delivery that can drift toward flat. Narration usually sits higher, character work lower — and it is the first dial to move when output sounds wrong.
Instant gets you a working clone from minutes of audio with no training wait: right for prototyping, drafts, and personal projects. Professional trains on much more source audio and takes correspondingly longer, in exchange for fidelity that holds up in commercial production. Many teams prototype on Instant and re-record the final with Professional.
Tag handling varies by model, and malformed tags fall through as text. Check the exact syntax — <break time=1s/>, self-closing — and use tags sparingly: for ordinary pacing, a period or an ellipsis is more reliable because pause-from-punctuation is native to every model.
Latency versus polish. eleven_turbo_v2 answers fastest, which matters for agents, apps, and anything conversational. eleven_multilingual_v2 is the produced-content choice: audiobooks, videos, localised material — where render time is irrelevant and quality is everything.
Not with capital letters — ALL CAPS tends to be read as spelling or yelling. Use the emphasis tag (<emphasis level=strong>) on the word, or restructure the sentence so the stress falls naturally; combined with a slightly lower stability, the model supplies emphasis on its own.
Open your assistant with this page preloaded as the source — great for follow-up questions like "which of these work in other apps?"