Eleven v4: the generation that pushes cloning fidelity far enough to reproduce bad training data

elevenlabs-v4

The new generation speech synthesis model from ElevenLabs, shipping in two variants: Eleven v4 for highest quality and Eleven v4 Turbo for real time use at a median inference latency of about 100 ms, with an official footnote stating that application and network latency are excluded. The vendor positions it above v3 on output quality, voice accuracy, consistency, emotion, delivery, audio tags and language coverage, and recommends migrating. We list it as the top closed source voice row in the audio domain because this generation changed the optimisation target from sounding better to sounding more like the source: timbre, cadence, delivery and other mannerisms are reproduced together, and the vendor concedes that accuracy and personal preference are not always the same thing and that you may still prefer how a voice sounded in v3. The engineering consequence is hard, since the burden of data cleaning moves back to the user. The showcase samples are explicitly labelled raw, with no EQ, compression, normalisation, de-essing or plosive removal, and the vendor says that if you hear clipping or plosives they are very likely in the training data and the model simply captured them accurately. What is deliverable comes down to three numbers: about 100 ms on Turbo; 10,000 characters per request, roughly ten minutes of audio, against 5,000 on v3 so long form text carries twice as many segment seams; and a coverage claim of 90 plus languages against a FAQ that enumerates 87, a gap we keep visible rather than smoothing away. Four boundaries matter. Cross language accent handling is a default behaviour change and not a toggle, so generating in a language that differs from the reference produces fluent native sounding speech in the target language, and the vendor calls a switchable version a research project with no timeline; any persona that depends on a native accent carried into a second language must be tested first. The control surface narrowed to Stability and Similarity, the Style and Speed sliders are gone, SSML is unsupported, and fine grained delivery moves to audio tags which the vendor admits are not perfect yet. Continuous model evolution is stated in writing, with training continuing after launch and behaviour possibly shifting over time, so teams treating a voice as a brand asset need periodic re-testing. Voice Design voices may also be less performative on v4. It is closed and not self-hostable, data must pass through ElevenLabs, and cloning compliance sits with the user; the self-hostable counterparts are CosyVoice 3, VibeVoice and Kokoro for speech and YuE2 for music under a non-commercial weight licence. This entry and the Eleven v3 already catalogued here are two generations of the same commercial stack rather than a replacement. Graded C (vendor claim): no verifiable third party benchmark, and no blind listening test by us.

C

CONFIDENCE

Vendor Claim

Official model card or keynote only, no independent re-test

~100ms

KEY METRIC

v4 Turbo 中位推理延迟(模型侧,不含应用与网络)

Vendor Claim · 2026-09

MATURITY

Production

research → demo → product → production

Our take

We grade it C (vendor claim). ElevenLabs publishes no verifiable third party benchmark: there is no MOS figure, no speaker similarity number and no WER we can cite, only the vendor own generational comparison plus a set of showcase samples, and we have run no blind listening test of our own. Contract section 13.5 therefore forbids an upgrade to A. What makes this entry credible within that grade is something else: the vendor labels its showcase samples as raw, with no EQ, compression, normalisation, de-essing or plosive removal, and states outright that if you hear clipping, plosives, uneven EQ or volume jumps then those artefacts are very likely in the training data and v4 simply captured them accurately. Publishing the flaws of your own output for anyone to hear is unusual in an audio model release, and it is why we list this as the top closed source voice row rather than merely relaying the marketing claim.

The case for the top of the ladder is a generational step on the cloning fidelity axis, paired with a real time latency that is rare at this quality tier. The change from v3 to v4 is not that it sounds better but that it sounds more like the source: timbre, cadence, delivery and other mannerisms are reproduced together, and the vendor itself concedes that accuracy and personal preference are not always the same thing and that you may still prefer how a voice sounded in v3. That concession amounts to admitting the optimisation target of this generation is fidelity rather than pleasing the ear. The engineering consequence is hard: the burden of data cleaning moves back to the user, because reference recording quality now determines output quality directly instead of being masked by model side beautification, and the vendor changed its training data guidance accordingly (stick to a single speaking style, with an explicit note that the guidance may change). What is actually deliverable comes down to three numbers: median inference latency around 100 ms on the Turbo variant, with an official footnote stating that application and network latency are excluded; 10,000 characters per request, roughly ten minutes of audio; and a language claim of 90 plus against a FAQ that enumerates 87. For contrast, v3 allows 5,000 characters per request so long form text carries twice as many segment seams, and Flash v2.5 is faster at about 75 ms but covers only 32 languages with text normalisation off by default. We keep the gap between 90 plus and 87 visible rather than smoothing it, because a headline coverage claim and an enumerable list have always been two different numbers.

Four boundaries must be understood before migrating. Cross language accent handling is a default behaviour change, not a toggle: when the generated language differs from the reference voice language, v4 produces fluent native sounding speech in the target language instead of carrying the reference accent across, and the vendor says making it switchable is still a research project with no timeline. If a product persona depends precisely on a speaker carrying a native accent into a second language, as in character voiceover or interview reconstruction, this upgrade silently changes the output and must be tested or avoided by staying on v3. The control surface narrowed: only Stability and Similarity remain, the Style and Speed sliders are gone, SSML is unsupported, and fine grained delivery moves to inline audio tags which the vendor admits are not perfect yet and remain an active area of investment. For a pipeline built on SSML or on the Style slider this is migration cost, not upgrade benefit. Continuous model evolution is stated in writing: training continues after launch, behaviour may shift over time and periodic re-testing is recommended, so any product that treats a voice as a brand asset should put that on an operations calendar because reproducibility of output is not under your control. In addition Voice Design voices may be less performative on v4 in the vendor own words, so projects built on Voice Design should compare both generations before moving.

Then the one that cannot be routed around: closed and not self-hostable, data must pass through ElevenLabs, and the compliance burden for voice cloning sits with the user because capability is not authorisation. For teams whose data cannot leave their own infrastructure, or who need to hold the voice weights themselves, the self-hostable counterparts in our audio domain are CosyVoice 3, VibeVoice and Kokoro for speech, and YuE2 for music with the caveat that its weights are CC-BY-NC-4.0 and therefore non-commercial. One selection note: v4 and the Eleven v3 already catalogued on this site are two generations of the same commercial stack, not a replacement. The vendor recommends migrating while also conceding that a few edge cases fit v3 better, so two coexisting rows on the ladder is the correct reading.