ElevenLabs API playground
This tool calls the ElevenLabs text-to-speech API directly from your browser with your own key, so you can audition voices and fine-tune voice settings on real text without code. Paste a voice ID, adjust stability, similarity boost, and style, and generate — the MP3 plays inline and downloads with one click.
How it works
On generate, the playground POSTs to
https://api.elevenlabs.io/v1/text-to-speech/{voiceId} with an xi-api-key
header, your text, and a voice_settings object carrying stability, similarity
boost, and style. ElevenLabs streams back MP3 audio; the browser wraps the bytes
in an object URL for an inline player and download link. There is no backend and
nothing is stored.
What each voice setting actually controls
Understanding the three sliders helps you get a useful take on the first try rather than burning characters through trial and error:
Stability (0–1) controls how deterministic each generation is. At 1.0 the model produces nearly the same delivery every time you generate the same text, which suits audiobooks and training data where consistency matters. At 0.0 the output is highly variable — each take sounds different, which can be useful for finding an expressive read but is unreliable for production. Most narration work lands between 0.35 and 0.65.
Similarity boost (0–1) controls how tightly the generated voice is tied to the original voice sample in the clone. High values make the voice sound more like the reference recording but also amplify any noise or room tone present in that recording. If a cloned voice sounds hissy or has a room echo, try lowering similarity boost before re-recording.
Style (0–1) amplifies the expressive style embedded in the original voice sample. At zero the voice reads in a relatively neutral, even way. Raising it adds the characteristic emotional colouring of the source voice — but it also slows generation and makes the output less consistent. Reach for style last, after stability and similarity boost are set.
Workflow: audition before you commit
Because ElevenLabs bills per character, a good workflow is to write a representative 50–100 character sentence — not your full script — and iterate on the voice settings with that short passage. Once the tone feels right, generate the longer content. This approach typically saves significant character usage when finding the right voice for a project.
Tips and notes
- Tune stability per use. Audiobooks want higher stability for consistency; characters and ads benefit from lower stability and more style.
- Similarity boost ≠ stability. Boost pulls toward the source timbre; raise it if the clone sounds generic, lower it if it sounds strained.
- Keep text natural. ElevenLabs reads punctuation and casing as performance cues — write the way it should be spoken. Dashes and ellipses produce pauses; exclamation marks raise energy.
- Watch the character count. Usage is per character, so trimming reduces both cost and generation time. The playground shows the count before you generate.
- Voice ID location. Find the voice ID in the ElevenLabs VoiceLab URL or on the voice’s detail panel. The ID is the alphanumeric string after
/voice/in the URL.