Eleven v4 Review: I Made a Short Film With It. Here's What It Gets Right, and What It Doesn't

I made a heist thriller, under two minutes long, where you never see anyone speak. "Pick Up, Kevin" is six increasingly desperate voicemails from a man named Victor, one phone call, and one last voicemail after the title card. Every line was voiced with Eleven v4, the text-to-speech model ElevenLabs released on September 28, 2026.

That film is the test this review is built on. Not a demo sentence, but a story where the voices carry everything. If the acting fell flat, the film fell flat.

Below is what I found, what it costs, and the bigger question you can't avoid once you hear it: How does a voice model this good impact voice actors/actresses?

The short version

Eleven v4 is the most convincing AI voice actor I've used, but the built-in sound effects aren't there yet, and you should budget credits for retakes.

What Eleven v4 is

Eleven v4 is ElevenLabs' new flagship voice model, built on a new architecture that reads the whole scene before it speaks a line. It shipped alongside a faster sibling, Eleven v4 Turbo, on September 28, 2026 (ElevenLabs).

Sources: ElevenLabs models docs, Eleven v4 page, best practices.

On the quality claims: Artificial Analysis, an independent leaderboard where people vote blind between voices, ranked v4 first at launch with an Elo of 1319, ahead of Cartesia Sonic 3.6 at 1276 (Tech Times). ElevenLabs also says about 75% of listeners preferred v4 in blind tests, but that's the company's own study, so weigh it accordingly.

How I tested it

I voiced a full short film on the Starter plan, using standard Eleven v4 for every line (not Turbo).

The cast was two voices. Victor is an American man in his early 30s who gets more frantic with each voicemail. Ruth is a British woman in her early 70s, and the twist of the film depends on her sounding exactly like who she seems to be. Neither is a real person: I created both with ElevenLabs Voice Design from a one-sentence description.

My process:

1. Wrote the script and locked it before generating anything
2. Generated several takes of every line and picked the keeper by ear. The keeper was usually the third or fourth take
3. Made every background sound (alarm, footsteps, helicopter, sirens, rain, a phone vibration, a police scanner) separately with ElevenLabs Sound Effects, plus a voicemail beep before each message
4. Fixed pauses and timing in the edit
5. Built the visual cutaways in Higgsfield, with no faces delivering lines and no lip-sync, then cut everything in CapCut

What Eleven v4 gets right

It acts instead of reading. Older text-to-speech reads a sentence. v4 reads the situation: ElevenLabs says it works out tone, pacing and character from the surrounding text before it speaks (ElevenLabs). In practice, Victor's panic builds across six voicemails the way a script would ask an actor to play it. Play the first voicemail ("Kevin. I'm inside the Halden Building.") and then the fourth ("Why is there a helicopter, Kevin?!"): same voice, a very different man.

The voice stays the same person. My keepers came from different rounds of takes, and Victor still sounds like one man from the first message to the last. That matches ElevenLabs' claim that regenerating a line no longer makes the voice drift.

Age and accent hold up. Ruth is the hardest part of the film to get right. An older British woman has to sound warm and harmless until the twist, and v4 gave me takes that did that without tipping into caricature.

You can cast from a description. Neither voice is a stock voice or a clone. I wrote one sentence for each in ElevenLabs Voice Design. For Ruth it was "a woman in her early seventies with a soft British accent... like a grandmother who has seen everything and is never surprised", and v4 played her as a character, not a reading.

It handles long scripts. One generation can run to 10,000 characters, about 10 minutes, double v3's limit (ElevenLabs docs). That matters if you narrate YouTube videos or audiobooks.

The language list finally reaches East Africa properly. Swahili was already supported in v3. v4 adds Amharic, Kamba, Luganda, Shona, Wolof and Zulu, and ElevenLabs says a voice now speaks each language with a native accent instead of drifting back to its original one.

Where Eleven v4 falls short

The built-in sound effects didn't work for me. ElevenLabs pitches tags like [light rain] or [phone buzzing] as a way to put sound design straight into the script. In my takes, the background sounds didn't come through. The one that did, a tire screech inside Ruth's line, sounded fake, so I replaced it with a separately generated screech. Missed tags were a known complaint with v3 as well (Unite.AI), and ElevenLabs' own docs say a tag the voice wasn't trained on may not come out well (best practices).

You pay for retakes. Tags are direction, not a guarantee. My keeper was usually the third or fourth take, and every regeneration spends credits. Plan for two to four takes per important line.

Pacing moved out of your hands. There's no speed slider and v4 ignores SSML break tags, so old v3 scripts with breaks in them need rewriting. I fixed every pause in the edit instead of in the generation.

The speed numbers need context. Turbo's ~150 ms figure is ElevenLabs' own test with network delay removed. A real voice agent also waits on transcription and the language model, so callers won't hear 150 ms end to end (Tech Times).

Eleven v4 vs v4 Turbo

Use v4 when the recording is the product; use Turbo when someone is waiting for the voice to answer.

Sources: Eleven v4 page, ElevenLabs models docs.

For a film, a YouTube voiceover or an audiobook, standard v4 is the one you want.

Pros and cons

Pros

• Delivery that sounds performed, not read, with emotion that builds across a scene
• The voice stays consistent across retakes, dialogue and long narration
• 90+ languages with native accents, including more African languages than v3
• 10,000 characters per generation, double v3
• Instant voice clones from about 10 seconds of audio; Professional Voice Clones supported
• Ranked first by an independent blind-listening leaderboard at launch
• Available on the free plan

Cons

• Inline sound-effect tags were unreliable in my testing
• Tags are hit and miss, so retakes eat credits
• No speed slider and no SSML breaks; pacing gets fixed in the edit
• Old v3 scripts need rewriting
• Turbo's headline speed doesn't include the rest of a voice agent's pipeline

Pricing

You can try v4 free; anything you publish for money needs a paid plan, which starts at $6 a month.

Source: Eleven v4 page, checked October 5, 2026. Prices change, so confirm on the live pricing page before you sign up.

What it cost me: every voice line and sound effect in the film was made on the $6 Starter plan.

What v4 means for the entertainment industry

v4 doesn't replace the star performance yet, but it makes the cheap end of voice work, and solo productions like mine, far easier, and that shift is already costing some voice actors their income.

Audiobooks. ElevenLabs' AI recreation of Michael Caine narrates a 13-hour audiobook of The Odyssey. Critics praised the voice but said the emotional scenes fell flat, which is exactly the gap v4 is built to close (Marketplace). So far it's a small business: AI audiobooks made about $729,000 in 2025, roughly 0.03% of the industry's revenue, mostly self-published titles that would never have paid for a human narrator.

Dubbing and localization. One voice can now speak 90+ languages with a native accent. For a creator, that means one video can reach Swahili, Spanish and Portuguese audiences without hiring three studios. For voice actors, dubbing is a big share of the work that could go away.

Games and animation. v4 handles character dialogue that reacts to the previous line, and Turbo makes characters that talk back in real time. Expect more AI-voiced side characters and NPCs, with human actors kept for leads.

Solo creators. My film had no cast and no booth, and every voice came from a $6-a-month plan. That's the biggest change: stories that used to need a production team can now be made by one person.

Voice actors. The small gigs are going first. Instagram content, YouTube explainers, online courses and budget audiobooks are increasingly voiced by AI, and one Indiana narrator told reporters her income was cut in half (LA Times via News Tribune). About 80% of voice work is freelance and outside union rules, according to the National Association of Voice Actors.

New protections. SAG-AFTRA's June 2026 agreement requires studios to notify actors before making a voice replica, get their sign-off, pay them the same as an in-person session, pay residuals, and never use replicas during a strike. The union is also pushing the No Fakes Act, which would make unauthorized use of anyone's voice illegal nationwide.

Licensing as income. ElevenLabs says more than 10,000 people earn from listing their voice in its library, with $22 million paid out. The rate is small, though: about $0.03 per 1,000 characters, around $1.20 per hour of use, and performers don't always know which projects their voice ends up in (Marketplace).

The lawsuits. In May 2026, journalists, broadcasters, podcasters and voice actors filed a class action in Illinois, Amer v. Eleven Labs, alleging their voices were collected and used for training without consent. The case is pending and isn't about v4 specifically (Tech Times).

Same tool, two outcomes

Whether v4 is good or bad for people depends almost entirely on how it's used: consent, credit and disclosure are what separate the two.

The rules I follow:

• Only use voices I have the rights to. Both voices in my film were designed from scratch, so no real person's voice was copied
• Say it's AI. Wherever I post Pick Up, Kevin, the caption says it was made with ElevenLabs and Higgsfield
• Read the license on any library voice before using it commercially
• When the performance is the whole product and there's a budget, hire a person, or pay the person whose voice you license

Final verdict

If you make anything where a voice has to act, Eleven v4 is the best tool available right now. My score: 8/10.

It carried a whole film for me: two characters, a twist and a slow build of panic, with no actors and no booth. What it won't do yet is design your soundscape inside the voice track, or nail a line on the first try every time. Make your sound effects separately and budget a few takes per line.

Start on the free plan to hear it with your own script. If you're publishing for money, move to a paid plan for the commercial rights.

FAQ