AI voices have been good for a while. Now they are getting directable. ElevenLabs Eleven v4, launched September 28, 2026 (with an updated announcement on October 1), is the company’s biggest text-to-speech release yet: a model that does not just read your script aloud — it performs it, taking stage directions written right inside the text. Alongside it came Eleven v4 Turbo, a low-latency twin built for live voice agents. Here is what v4 actually changes, what it costs, and whether the hype survives contact with reality.
What Is ElevenLabs Eleven v4?
ElevenLabs is the London-based voice AI startup founded in 2022 that became the default tool for synthetic narration — from YouTube voiceovers to audiobooks to game dialogue. Eleven v4 is its new flagship speech model, built on a new architecture and available across the company’s ElevenCreative studio, ElevenAgents voice-agent platform, and the ElevenAPI.
The headline idea: earlier generations of text-to-speech mapped words to audio. Eleven v4 interprets how the words should be spoken first — tone, pacing, emotion, character, context — and then synthesizes speech that carries that performance. According to ElevenLabs, the model can sound dramatic, tender, urgent, comedic, or conversational from the same voice, and keeps each speaker’s identity consistent across long passages and regenerated lines. Independent testing seems to agree: on launch day, the benchmarking firm Artificial Analysis placed v4 #1 on its Speech Arena leaderboard, with blind listeners preferring it roughly three times out of four. (ElevenLabs’ own tests put that preference share at about 75% — the company’s figure, from its own measurements, not a third party’s.)
The Big Feature: Inline Emotion Tags
The most practical change is control. With v4 you can direct delivery from inside the script using inline audio tags for emotion, pacing, accents, laughter, and sound effects — no re-recording, no separate audio pass:
- [laughs] — natural laughter woven into the line, not pasted on top.
- [speaks angrily with a British accent] — emotion and accent in one tag.
- [light rain sound], [cellphone vibration] — ambient sound effects generated alongside the voice.
You can also write directions in plain English (“say this like breaking news”) and use IPA phonetic spellings for tricky names and terms, which v4 handles more reliably than its predecessor. This is the real workflow win for creators: audition a restrained reading and a dramatic one against the same visuals, pick the one that serves the scene, and move on. The caveat is that tag-following is imperfect by ElevenLabs’ own admission — an over-performed line in a customer-support or education context will feel wrong, so test and listen before shipping.
Eleven v4 Turbo: Expressive Speech at Agent Speed
The eternal voice-AI trade-off has been quality versus latency: the best-sounding voices answered slowly, and the fast ones sounded flat. v4 Turbo is ElevenLabs’ answer — the same expressive range, optimized for streaming conversations. The company reports a median inference latency of about 100 ms and a median time to first speech of about 150 ms over WebSocket streaming.
One honest detail worth knowing: that 150 ms figure was measured with network latency removed, in a September 2026 benchmark. It is the right way to compare models, but the wrong way to promise an experience — real deployments add network transit, the language model behind the agent, audio buffering, and playback. Still, sub-second spoken responses are now realistic, which changes what voice agents can do in support, sales, and phone automation. For an in-depth look at how voice agents are evolving, our Tab AI assistant breakdown covers another angle of the trend.
90+ Languages and 10-Second Voice Cloning
Language coverage jumps from 70+ languages in v3 to 90+ in v4 — including Cantonese, Mongolian, and Odia — with ElevenLabs singling out Japanese, Brazilian Portuguese, Mandarin, and Cantonese as the biggest quality gains. A cloned voice can also switch languages while keeping its original identity and taking on the target language’s accent naturally: a Spanish speaker’s clone reading German still sounds like that person, with a native German accent.
Voice cloning got a serious upgrade too. Instant Voice Clones now need just 10 seconds of reference audio, and speaker identity holds together better across long texts — a weak point of v3, where voices sometimes drifted over the course of a chapter. Professional Voice Clones return in v4 (they were missing from v3), and each request can now hold up to 10,000 characters — roughly ten minutes of audio — with more reliable request stitching for audiobooks and long narration. Every clone requires the voice owner’s verified consent, and ElevenLabs blocks cloning of celebrities and other high-risk voices.
Eleven v4 vs v4 Turbo: Which to Use?
| Eleven v4 | Eleven v4 Turbo | |
|---|---|---|
| Best for | Audiobooks, ads, narration, character dialogue | Voice agents, support calls, real-time chat |
| Latency | Quality-first | ~100 ms inference / ~150 ms first speech |
| Languages | 90+ | 90+ |
| Cloning | 10-sec instant + professional clones | Same |
| API list price | $80 / 1M characters | $40 / 1M characters |
ElevenLabs v4 Pricing
ElevenLabs ran a two-week launch offer — 72% off API pricing, at $22/1M characters for v4 and $11/1M for Turbo — but that discount ended October 12, 2026. For planning purposes, use the list rates: $80 per million characters for v4 and $40 for Turbo. (Character-based pricing: roughly 15–16 words of English equal ~100 characters, so a 10-minute narration costs a few dollars in API terms.)
The subscription tiers start with a Free plan (10,000 credits a month, about 10 minutes of speech; no commercial license, attribution required), then Starter at $6/month, Creator at $22, Pro at $99, Scale at $299, and Business at $990, with annual billing at ten months’ cost. Any paid plan includes a commercial license, so monetized YouTube videos and ads are covered from Starter up. A creative-suite launch bonus gives Creator plans and above free v4 usage up to double their monthly credits. Always check the official ElevenLabs v4 announcement for current pricing before committing — this space moves fast.
Pros and Cons
Pros:
- Ranked #1 on Artificial Analysis’ blind-listener leaderboard on launch day.
- Inline emotion tags give line-level direction without re-recording.
- 90+ languages with real accent adherence when switching languages.
- 10-second instant cloning plus restored professional clones.
- Turbo model brings expressiveness to real-time agents (~150 ms first speech).
- Consent verification and high-risk voice blocking for cloning.
Cons:
- API list pricing is steep versus Google’s Gemini TTS (~$16.50/1M characters) — though Google offers no voice cloning.
- Tag-following is imperfect; dramatic delivery can feel wrong in serious contexts.
- Long works still need section-by-section generation and stitching.
- The 72%-off launch discount has expired; production costs are the list rates.
Who Is Eleven v4 For?
Content creators (YouTubers, podcasters, course makers) get the biggest immediate win: directed, multilingual narration without a recording studio. Pair it with AI video and you have a full production line — see our guides to the best AI video generators and Vidu Q4’s photo-to-cinema model for the visuals. Game studios and audiobook publishers gain character dialogue that reacts to what was just said, with voices that stay consistent over long scripts. Voice-agent developers finally get expressive speech at agent-grade latency — ElevenLabs jointly optimized Turbo with its ElevenAgents platform. If your needs are simpler, note that Suno’s speech beta covers the spoken-word-plus-music angle for creative audio.
Verdict: The New Benchmark, If You Can Afford It
ElevenLabs Eleven v4 earns its launch-day #1. The shift from reading to performing is real: emotion tags change the creator workflow the way color grading changed video editing — a new layer of craft on top of a solved problem. The honest trade-offs are price and imperfection: it costs more than Google’s alternatives, tags don’t always land, and the launch discount is gone. But for anyone whose product is voice — audiobooks, dubbing, agents, ads — v4 is now the model to beat. Start with one script, A/B it against your current voice, and let your ears decide.
Frequently Asked Questions
Is Eleven v4 better than v3?
Yes, measurably. Independent blind-listener testing put v4 at #1 on the Artificial Analysis leaderboard on launch day, and ElevenLabs reports listeners preferred it roughly three out of four times. The practical upgrades: inline emotion tags, 90+ languages (up from 70+), 10-second instant clones, professional clones, and the low-latency Turbo twin.
How do Eleven v4 emotion tags work?
You insert tags like [laughs] or [speaks angrily with a British accent] directly into your script, and the model shapes the delivery — emotion, pacing, accent, even sound effects like [light rain] — around that line. You can also write plain-English directions (“read this like a weather alert”). Expect to regenerate a few lines; tag-following is good but not perfect.
Is ElevenLabs voice cloning safe and legal?
ElevenLabs requires verified consent from the voice’s owner for every clone, blocks celebrity and high-risk voices, and bars deceptive impersonation in its Prohibited Use Policy. Businesses using cloned or branded voices in agents must disclose that callers are talking to AI. Those are meaningful safeguards — but consent checks depend on uploaders telling the truth, so treat the technology itself with care.
How much does ElevenLabs cost in 2026?
Subscriptions run from Free ($0, ~10 minutes/month, no commercial license) through Starter ($6), Creator ($22), Pro ($99), Scale ($299), and Business ($990), with commercial licenses from any paid plan. API pricing at list rates is $80 per million characters for v4 and $40 for Turbo — the 72%-off launch offer ended October 12, 2026.
No comments:
Post a Comment