🎙️ ElevenLabs v3 Guide: Audio Tags, Best Settings & Realistic AI Voiceovers

Voice is 50% of a video's perceived quality — viewers forgive average visuals but click away from robotic narration in seconds. ElevenLabs v3 is what we use for every voiceover across our channels. This is the complete practical guide.

What makes v3 different

Older TTS read text. v3 performs it. The model understands emotional context and — critically — accepts audio tags: inline stage directions that shape delivery without being spoken.

[thoughtful] The strange thing is... nobody noticed for years.
[whispers] And when they finally did — [beat] it was too late.
[excited] But here's where the story turns!

Tags we use constantly:

• Emotion: [excited], [thoughtful], [sad], [nervous], [confident]
• Delivery: [whispers], [shouts], [slowly], [rushed]
• Non-verbal: [laughs], [sighs], [gasps], [clears throat]
• Pacing: [beat], [pause], ellipses ... for natural hesitation

Full walkthrough of the tag system:

Settings that actually matter

• Stability — the big one. High stability = consistent but flat; low stability = expressive but occasionally wild. For narration we run low-to-mid stability and regenerate the rare weird take. For character dialogue, go lower still.
• Similarity: keep high to stay true to the chosen voice.
• Model: use the v3 model for anything with emotional range. Older models are fine for utility audio only.

Rule of thumb: fix delivery in the script (tags), not in the sliders. Settings set the range; tags direct the performance.

The workflow we ship with

1. Write for the ear — short sentences, contractions, concrete words.
2. Tag the script while writing, not after: mark the emotional beats as you feel them.
3. Generate per paragraph, not one giant block. Energy stays fresh, retakes are cheap, and one bad take doesn't cost the whole render.
4. Keep one voice per channel. Voice = brand for faceless content. Save the voice and settings; never drift.
5. Master lightly: normalize loudness (target around -14 LUFS for YouTube), add a subtle music bed at -20dB or lower under narration.

Prompt-engineering deep dive — how phrasing and punctuation shape delivery:

Multi-speaker dialogue

v3 handles multi-speaker generation with natural turn-taking — the technology behind AI podcasts and character scenes. Assign distinct voices per speaker, keep lines short (real people interrupt and react), and use non-verbal tags ([laughs], [hmm]) for listener realism:

Cost control

Pricing is character-based, so waste = regeneration. Three habits keep our bills sane:

• Generate per-paragraph (retake only what failed)
• Proof the script before generating, aloud
• Draft new voices/styles on short test lines, not full scripts

For where the voiceover fits in the full production stack — script → voice → visuals → edit — see our faceless YouTube channel guide.