What does stability control in ElevenLabs?

Stability governs how consistent the voice sounds across a generation. Lower values make the delivery more varied and emotional but can wander or glitch. Higher values are steadier and more monotone. Most narration sits around 0.4 to 0.6.

What is similarity boost?

Similarity boost controls how closely the output matches the original voice sample. Higher values cling tightly to the source timbre, which helps cloned voices but can amplify any background noise present in the training audio.

What does style exaggeration do?

Style amplifies the speaking style and emotion of the original voice. It can add life to flat readings but raising it increases generation time and instability, so keep it low unless you specifically want a dramatic delivery.

What settings are best for audiobooks?

Audiobooks favor consistency over drama — try stability around 0.5 to 0.65, similarity boost around 0.75, and style near 0. This keeps a long read steady so the listener does not get jarred by sudden tonal shifts.

Should I enable speaker boost?

Speaker boost increases fidelity to the original speaker at a small latency cost. It is usually worth enabling for cloned voices and narration, and optional for high-throughput conversational use where latency matters.

ElevenLabs Voice Settings Advisor

ElevenLabs voice settings advisor

ElevenLabs gives you three sliders that make or break a generation: stability, similarity boost, and style exaggeration. The right values depend entirely on what you are making — a steady audiobook narrator wants the opposite settings from a wild animated character. This advisor maps your use case and tone to a recommended starting point.

How it works

Each slider trades off in a predictable way:

Stability (0–1) — low means expressive and variable but risky; high means consistent but flat. The sweet spot for most spoken content is the middle.
Similarity boost (0–1) — how tightly the output hugs the source voice. High fidelity to a clone, but it also reproduces source noise.
Style (0–1) — amplifies the original speaking style. Powerful but it slows generation and reduces stability, so use it deliberately.

The advisor starts from proven presets per use case, then nudges them based on whether you prioritize consistency or expressiveness.

Tips for dialing it in

Generate the same line twice at your chosen settings. If the two takes sound wildly different, raise stability.
Noise in, noise out. If a cloned voice sounds hissy, lower similarity boost rather than re-recording.
Reach for style last. Get stability and similarity right first; only add style if the read still feels lifeless.
Long-form needs steadiness. For anything over a minute, bias toward higher stability so the listener is not jolted by tonal swings.