ElevenLabs

ElevenLabs' current top music model, released in September 2026, improves prompt adherence and arrangement depth over Music v2. It keeps the same composition plans, audio reference and inpainting workflows, and composition-plan chunks take up to 6,132 characters.

Music v2, announced in May 2026, added genre switches within a single song, section-level inpainting and chunk-based composition plans for structured tracks. Music v2.5 has since superseded it, though ElevenLabs keeps v2 available.

Real-time variant of Eleven v3 built for voice agents, with about 280 ms latency, audio tags and 70+ languages. It debuted in ElevenAgents in February 2026 and has since been succeeded by Eleven v4 Turbo.

Scribe v2 is ElevenLabs' batch speech-to-text model for recorded audio in 90+ languages, with diarization of up to 32 speakers, word timestamps, entity detection and up to 1,000 keyterms. It replaces Scribe v1, slated for removal July 9, 2026.

Streaming counterpart to Scribe v2, transcribing live audio over a WebSocket with about 150 ms latency across 90+ languages. It offers voice activity detection, manual commit control and word timestamps, but no speaker diarization.

Generates sound effects from a text prompt, with clips of 0.5 to 30 seconds and a prompt-influence control. Version 2, released September 2025, added looping and longer clips; it is the API's default and only sound model.

Text to Dialogue endpoint running Eleven v3: a list of speaker turns, each with its own voice and audio tags, becomes one multi-speaker audio file. Public since August 2025; ElevenLabs now recommends Eleven v4 on this endpoint.

ElevenLabs' first music model, launched in August 2025, composes full songs, vocal or instrumental, from text prompts and gives section-level control over genre, structure and lyrics. The docs list it as outclassed by Music v2 and Music v2.5.

Eleven v3 brought inline audio tags such as [whispers] and [laughs], plus multi-speaker dialogue, to 70+ languages, with 5,000-character requests. It left alpha in February 2026 and is now previous generation, superseded by Eleven v4.

Eleven Flash v2 is the English-only model in ElevenLabs' low-latency Flash tier, with about 75 ms model latency and up to 30,000 characters per request. It replaced Turbo v2, which ElevenLabs now lists as deprecated.

Multilingual member of the Flash tier: 32 languages, about 75 ms model latency, 40,000-character requests and a 50% lower per-character API price. ElevenLabs lists it as its current fast, low-cost option and the replacement for the deprecated Turbo v2.5.

Turbo v2.5 extended the Turbo tier to 32 languages in July 2024, adding Hungarian, Norwegian and Vietnamese. ElevenLabs has since deprecated it and suggests Flash v2.5, which covers the same languages at lower latency.

ElevenLabs Voice Isolator strips background noise, music and ambient sound from a recording and returns the speech alone. It accepts audio or video files up to 500 MB and one hour, with API access since July 2024.

This speech-to-speech model powers Voice Changer, converting a recording into another voice in 29 languages while keeping whispers, laughs and other delivery details. It succeeded English STS v2 in February 2024 and remains the recommended Voice Changer model.

English-only Voice Changer model that re-voices an uploaded recording with a chosen voice while keeping its timing and emotion. It launched Speech to Speech in November 2023; ElevenLabs now says multilingual STS v2 often does better, even on English.

Eleven Turbo v2 was ElevenLabs' first English low-latency model, released in November 2023 with about 400 ms latency. It is now deprecated in favor of Flash v2, which ElevenLabs describes as functionally equivalent but faster on average.

A 29-language text-to-speech model from August 2023 that ElevenLabs documents as its most stable on long-form output, with 10,000-character requests. Now labelled previous generation behind Eleven v4, it is still recommended for narration and number normalization.