ElevenLabs launched Eleven v4 and v4 Turbo, and the benchmarks back the announcement: the model ranked first on Artificial Analysis’s independent text-to-speech evaluation, preferred by roughly 75 percent of listeners over competing models in blind head-to-head tests. For video editors and content creators, what actually matters isn’t the leaderboard placement — it’s what the model does differently and whether the workflow fits into a real production day. Here’s what changed and how to put it to use.
Image: ElevenLabs
What’s New in Eleven v4
Four changes matter for video work. The model now supports more than 90 languages (up from 70 in v3), which means a single project can cover more territory without switching tools. Voice cloning dropped from a longer sample requirement to 10 seconds of audio — a short clip, a phone recording, anything with a clear voice is now enough to generate a clone you can deploy across a full script. Speaker stability across regenerations is dramatically better: if you redo a line because the first take felt flat, the voice identity doesn’t drift between takes, which was a real problem in v3 for long-form narration. And Professional Voice Clones regain full emotional range, which v3 had partially limited.
ElevenLabs describes v4 as their “most emotive text-to-speech model ranked #1 by Artificial Analysis.” That’s a real improvement over v3, which required significant prompt engineering to get expressive delivery. V4’s expressiveness is more context-aware — the model reads the surrounding text and adjusts pacing and tone without you having to direct every sentence explicitly.
Choosing the Right Model: v4 vs. v4 Turbo
The rule is simple: if a human is waiting on the other end — a voice assistant, a real-time agent, a phone call — use v4 Turbo. It has a median inference latency of around 100ms and time to first speech of about 150ms. If you’re rendering a voiceover file where nobody is waiting in real time, use the full v4 for the best quality. Both are available via the API and ElevenLabs’ own dashboard; specify eleven_v4 in your model ID if you’re working programmatically.
Step 1: Your First AI Voiceover
Go to elevenlabs.io and open Text to Speech. The right starting point is a voice from ElevenLabs’ library — there are hundreds of options sorted by style (narrative, conversational, energetic) and language. Pick one that matches your content’s register. Paste a 30-to-60-second script written for listening, not reading: shorter sentences, contractions, natural pauses. Generate one baseline version before adjusting any settings.
Listen to the full output before deciding anything needs fixing. Most of what sounds off on a first pass is in the script, not the model — rewrite the awkward sentence, then regenerate. Save the generation, along with the voice, model, and script you used. With v4’s speaker stability, your next generation of the same script will sound consistent with the first, so you’re not rebuilding from scratch every time you need to revise a line.
Voice Cloning in 10 Seconds
An Instant Voice Clone takes a short audio sample — 10 seconds is enough — and generates a clone you can run scripts through immediately. The quality scales with the recording clarity: a clean phone recording in a quiet room will produce better results than a compressed clip with background noise. The clone isn’t perfect for every sentence on the first pass, but it gives you something you can actually use in a draft and iterate on.
Professional Voice Clones (available on higher tiers) use more data and return a better result with full emotional range. If you’re delivering narration for a client who needs their own voice maintained across a series of videos, a PVC is worth setting up properly. For a quick turnaround where a close approximation works, the Instant Clone gets you there in seconds.
Directing the Delivery: Inline Audio Tags
V4 supports inline audio tags inside your script text. These are bracketed instructions that tell the model how to deliver the line: [whispers], [laughs], [excited], [sighs], [crowd applause]. Place them inline, directly in the text you’re generating from, and the model follows them in sequence. V4 follows multi-tag sequences more reliably than v3 did, which means you can actually direct a performance without rewriting the text around it.
Note: v4 disabled the SSML <break> tag from earlier versions. Use [pause] instead for natural stops. The inline tag system is more flexible than SSML for expressive work, though it requires a bit of experimentation to find the tag phrasing the model responds to most consistently for your particular voice and use case.
Multi-Speaker Scenes and Longer Production
For two-person narration, tutorials with a host and a guest voice, or explainer videos where different sections have different presenters, the Dialogue mode handles multi-speaker scripts in a structured format — you assign lines to voices, specify the pacing, and get back a single mixed file. ElevenCreative Studio, the platform’s timeline-based tool, takes this further: you describe the project in plain language (length, tone, structure, transitions), and the Studio Agent drafts a first cut with clips, voiceovers, sound effects, and captions already in place. It’s best for projects where you need a fast, structured starting point and will refine from there, not for high-control long-form narration where you want to manage every line yourself.
Dubbing and Localization
The Dubbing tool takes an existing video or audio file, translates and re-voices it into another language, and preserves the original speaker’s voice characteristics in the target language. With v4’s 90-language support, a single English-language piece can be adapted into 30 or 40 language versions in the same session. The quality varies by language — Japanese, Brazilian Portuguese, Mandarin, and Cantonese are specifically called out as seeing the biggest improvement in v4. For high-priority markets where the output needs to hold up to native speaker scrutiny, do a quality check before delivery.
Music v2.5 for Scored Content
ElevenLabs added API support for Music v2.5 in this release cycle, available via the music_v2_5 endpoint. The music generation is separate from the voiceover workflow — you’re generating a score or a music bed rather than a narration — but the two can run in sequence through the same account and API. This makes sense for content where the voice and the music are separate production elements that need to be kept independent for final mix reasons, rather than collapsed together in a single output the way Suno Speech approaches the problem. Which model approach fits your project depends on how much post-production control you need over the audio mix.
What It Costs
The API is priced at $22 per million characters for Eleven v4 and $11 per million characters for Eleven v4 Turbo. A typical 60-second narration script runs around 800-1,000 characters, which means roughly $0.02 per generation at v4 rates. The free tier gives about 10,000 credits per month — enough to experiment with several projects before needing a paid plan. ElevenCreative is available to Creator+ subscribers, who get up to 2x monthly credits for it without additional charges.
Bringing It Into Your Edit
Export your generation from ElevenLabs as an MP3 or a WAV (higher quality plans give you the WAV option). Import it into Premiere Pro, DaVinci Resolve, Final Cut Pro, or CapCut the same way you’d import any audio file — drag it to a new audio track, line it up with your picture, and mix. In Resolve, the Fairlight page gives you full EQ and dynamics processing; in Premiere, Essential Sound handles most of what a narration mix needs. Set a high-pass filter on the voice track (typically around 80Hz) to clean up low-end rumble, add a touch of compression to even out level differences between sentences, and balance against your music bed and room ambience.
The AI voice is clean by definition — it doesn’t have the room character, microphone coloration, or natural variation of a real recording, which means it can sound slightly disconnected when cut next to real-camera footage. If the project mixes AI narration with real-world audio, a light reverb matched to the project’s acoustic environment closes that gap fast. The same principle applies to video: AI-generated B-roll next to camera footage benefits from a matching grain pass to unify the texture — adding a real film grain overlay in Premiere, Resolve, or Final Cut is a fast way to make synthetic footage feel consistent with real-camera footage in the same sequence. The voice equivalent is a touch of room character that makes the voiceover feel like it belongs to the same physical space as everything else in the mix.
Sources:
TechCrunch — ElevenLabs’ new v4 speech model supports more expression control and 90 languages
ElevenLabs — Eleven v4 (official announcement)
ExplainX.ai — Eleven v4 & v4 Turbo Launch
ElevenLabs — ElevenCreative Studio overview (official docs)
NovoAds — Eleven v4: What ElevenLabs’ New Voice Model Changes for AI Ad Voiceovers
