Skip to content
voicefieldguide

Eleven v4 vs ElevenLabs v3 for voiceovers

Updated Invalid Date

The picks

Press play to hear each voice read this guide's sample line.

These picks compare my fifteen voices. I own them, and ElevenLabs may pay me when you use them.

Titan Epic·Performance comparison

Its fiction and trailer fit gives an expressive line room to move. Plain instructions would sound overacted.

0:15

Heisenberg·Long narration

An even long-form voice exposes continuity between paragraphs. It needs room in a short edit.

0:12

Derek·Names and numbers

Broadcast delivery is a useful casting choice for factual test copy. The formal tone can stiffen casual scripts.

0:12
Add Titan Epic to My Voices

Pick Eleven v4 for voiceovers. ElevenLabs calls it a net upgrade over ElevenLabs v3 with “better results in almost every case”, and it doubles the characters you can send in one API request, from 5,000 to 10,000. For a Professional Voice Clone the choice is already made. ElevenLabs’ v4 launch page says PVCs weren’t supported in v3 and are back in v4. All fifteen of my voices are PVCs, so mine live on v4. Below is what changed, what came along unchanged, and the one narrow case where I’d still compare the two.

What changed between ElevenLabs v3 and Eleven v4?

Eleven v3 Eleven v4
Model ID eleven_v3 eleven_v4, plus eleven_v4_turbo for real-time use
Announced June 3, 2025 September 28, 2026
Languages 70+ 90+
Characters per API request 5,000 10,000
Professional Voice Clones not supported supported
Audio tags yes yes, followed more accurately
Voice settings Stability (the product guide lists Speed, Similarity and Speaker Boost as unavailable on v3) Stability and Similarity
SSML and break tags no no

ElevenLabs’ models page translates those character limits into roughly five and ten minutes of audio per request. On long-form work that halves the number of pieces you stitch together, which is reason enough on its own.

Can I use a Professional Voice Clone on ElevenLabs v3?

You shouldn’t, and ElevenLabs has said so twice. At the June 2025 launch of v3, it warned that Professional Voice Clones weren’t fully optimized for the model and suggested an Instant Voice Clone or a designed voice for v3 projects. The v4 launch page is blunter. It simply says PVCs weren’t supported in v3. ElevenLabs’ best-practices guide now sends anyone using a PVC or a Voice Library voice to v4.

On v4, the voice owner trains the PVC for the new model from My Voices. According to ElevenLabs’ v4 docs, a plus button next to Eleven v4 starts that fine-tune, and PVC support reached accounts over the days after launch. You don’t have to do anything here. The step belongs to the owner. The article-specific clips on this page are rendered on v4. If some other library PVC sounds off on v4, check whether its owner has trained it for the model.

Which voice settings does Eleven v4 have?

Two. ElevenLabs’ v4 docs say Style and Speed sliders aren’t available and SSML isn’t supported. There’s one wrinkle. The general Text to Speech product guide still has a Speed FAQ claiming speed works on every model, so if your app shows a speed control on v4, the docs contradict each other. Go by what you hear.

  • Stability sets how consistent the delivery stays from one generation to the next. Lower values open up the emotional range, and higher values keep the read steady.
  • Similarity sets how closely the output sticks to the original voice. ElevenLabs warns that a high value on a voice built from poor audio can bring that audio’s noise and artifacts back.

In practice, length and pace on v4 live in the script. A read that runs long gets fewer words. Honestly, I think that’s the better habit anyway.

Do audio tags and break tags work the same on Eleven v4?

Audio tags carry over. ElevenLabs’ help center says the same tags work on both models and recommends v4. [whispers], [sighs] and [laughs] behave as they did on v3, and v4 follows them more reliably. My audio tags guide covers keeping them predictable.

Break tags still do nothing. Neither model supports <break>. If you came to v3 from Multilingual v2, you already rewrote those pauses as punctuation, and that work stands.

What sounds different on Eleven v4?

  • Other languages. When a voice speaks a language other than its own, v4 aims for fluent, natural speech in that language without carrying the original accent across. ElevenLabs says that’s deliberate. A v3 project that relied on an English accent bleeding into Spanish lines needs a test before it moves.
  • Designed voices. ElevenLabs says Voice Design voices work on v4 but may be less performative than on earlier models.
  • Pronunciation. For v4, ElevenLabs documents IPA wrapped in forward slashes, and phoneme tags aren’t needed.
  • Real time. eleven_v4_turbo is the low-latency variant for agents, with median inference latency around 100 milliseconds. Recorded voiceovers belong on standard v4.

When is ElevenLabs v3 still worth testing?

ElevenLabs wrote “almost every case”, and that leaves a sliver of room. Say you have a running series on v3 that uses an Instant clone or a designed voice, and the read already works. Generate the next episode’s opening on both models before you switch over. Listeners do notice when a narrator’s delivery changes halfway through a series. Anything new should start on v4.

How do I move an API project from v3 to v4?

  1. Set model_id to eleven_v4 explicitly. Leave it out and the text-to-speech endpoint defaults to eleven_multilingual_v2.
  2. Strip any SSML. v4 does not support it.
  3. Trim voice_settings. v4 uses stability and similarity, so style or speed values in the request won’t change the read.
  4. Re-test numbers and names. A new model is a new reader. Run your trickiest lines again before you batch anything.

Which of my voices gain the most from Eleven v4?

Titan Epic benefits most from the wider emotional range. It has rasp, lift and a big performance, made for a suspense beat or a trailer line, and a tag like [quietly] before a reveal gives it somewhere to climb from. Feed it plain information and it overacts. Keep it for scripts with real stakes. See Titan Epic for storytelling.

Heisenberg gets the most out of the 10,000-character API requests. It’s low and even, built for long-form education and narration, and a longer request means fewer seams in a chapter. Its calm, spaced delivery suits long reads better than short ones, though, so I’d rarely use it for anything under a minute. See Heisenberg for audiobooks and long-form.

I’d compare performance with Titan Epic, paragraph continuity with Heisenberg, and factual copy with Derek. These clips are v4 examples; they aren’t a controlled v3 comparison.

Related guides

Other jobs where I use the same voices.

Nothing playing

Press play on any voice

↑ ↓ moveEnter openShift Enter playEsc