ElevenLabs launches Eleven v4, its most expressive text-to-speech model

Background

Text-to-speech (TTS) systems turn written text into spoken audio. For a long time the main goal was intelligibility: making sure every word was pronounced clearly and in the right order. Once that became reliable, attention moved to prosody — the rhythm, stress and intonation that carry meaning beyond the words themselves. The same sentence can be a joke, a warning or a comfort depending on how it is delivered, and a system that reads everything in one flat register loses that layer. A second, separate concern is latency: how long a system takes to produce audio after receiving text. For pre-recorded uses such as audiobooks, a delay of a second or two hardly matters; for a voice assistant that has to answer in real time, it matters a great deal.

What was announced

On 28 September 2026, ElevenLabs launched Eleven v4 and described it as its most emotive text-to-speech model yet^[1]. A low-latency variant, Eleven v4 Turbo, launched alongside it^[1]. The two versions follow a pattern common in speech systems: a full model aimed at quality, and a faster companion aimed at settings where response time is the binding constraint, such as an interactive agent or a live conversational interface.

Key details

According to the company, the model was designed to interpret tone, pacing, emotion, character and context^[1]. That list points at a shift in what the system is asked to do. Tone and pacing concern how a line is delivered; character concerns who is speaking; context concerns the surrounding situation. The announcement argues that a single line of text should sound different depending on who is saying it and to whom^[1] — that is, the same words from an adult to a child, or from one colleague to another, ought not to come out identical.

ElevenLabs also says the model was ranked first by Artificial Analysis^[1], a benchmarking organisation that publishes comparisons of AI models. Rankings of this kind usually aggregate automated evaluations, and they measure only what the tests measure.

The company further says that about 75% of listeners preferred Eleven v4 over competing models in blind head-to-head tests^[1]. In a blind comparison, listeners hear samples without being told which system produced which, and choose the one they prefer; a figure of roughly three in four means the model won close to three quarters of the paired comparisons in which it took part. Subjective tests like this capture something automated metrics do not, since intelligibility scores cannot say whether a voice sounded warm, bored or sarcastic. They are also sensitive to how the samples were chosen and which systems were included, and a company reporting its own comparison is describing a result it commissioned or ran.

Why it matters

Speech is used far beyond reading text aloud. It narrates audiobooks, dubs video into other languages, gives voice to assistants and screen readers, and increasingly serves as the output layer of conversational AI systems. In most of these settings, naturalness is not decoration: a voice that misplaces emphasis can make a correct answer sound uncertain, or a friendly one sound curt. Work on expressiveness aims at that gap.

Latency is the other half of the problem, and it is the reason a Turbo variant ships at the same time. A model that sounds excellent but takes too long to begin speaking is hard to use in a conversation, where pauses are themselves meaningful. Announcing both versions together is a way of saying the technology is meant for live use as well as recorded use.

The broader significance lies in where the effort is going. Earlier generations of speech synthesis competed mainly on whether the output sounded human at all. Eleven v4 is presented as competing on interpretation — on reading a line the way a person would, taking account of speaker and audience. Whether the preference numbers hold up under independent testing is the kind of question that only later evaluations answer, and the company's own claims should be read as claims.