ElevenLabs Music v2 and Dubbing v2: The June Audio Drops, Decoded

Between May and August 2026, ElevenLabs rebuilt a large chunk of its audio stack. Music v2 landed in the API on June 15 with a fundamentally different way to control a generated track. Dubbing v2 — which conditions on the original speaker's performance instead of just the transcript — launched May 28 and got its own API on August 6. And on July 9, the company pulled the plug on three legacy models: eleven_monolingual_v1, eleven_multilingual_v1 and scribe_v1.

None of these are Versely-hosted models, so this is a decode-and-translate post: what actually changed inside ElevenLabs, and how to hit the same creative briefs with the audio stack Versely does carry.

Music v2: composition plans instead of a single prompt

The old ElevenLabs music flow (music_v1) took a single text prompt in and produced one loosely-shaped track out. Music v2, which became generally available in the API on June 15, 2026, replaces that with what the changelog calls "chunk-based composition plans" — built from GenerationChunk and AudioRefChunk segments that give "finer control over structure, pacing and arrangement than the prompt-based music_v1 flow." In practice that means specifying a track as a sequence of labeled sections — intro, build, drop, outro — rather than hoping one paragraph of description produces the arrangement you're picturing. The two chunk shapes stay backward compatible, so a team can migrate one endpoint at a time; MusicPrompt still works for anyone staying on v1.

That's a genuinely different mental model from most prompt-driven music tools, and it's worth knowing even outside ElevenLabs' own customer base, because it previews where the category is heading generally — from "describe a vibe" toward "describe a structure."

Dubbing v2: dubbing that keeps the performance

ElevenLabs' dubbing model got the same kind of rebuild. Traditional AI dubbing transcribes, translates, and re-voices from the translated text — which is a lot of why so many dubs sound flatter than the original. The translated sentence is correct, but the delivery is generic: a neutral text-to-speech read has no way to know where the original speaker paused for effect, raised their voice, or rushed a line, because none of that survives a text-only pipeline. Dubbing v2, which ElevenLabs introduced on May 28, 2026, conditions directly on the original audio instead, carrying "tone, pacing, delivery, and emotional intent" across the translation rather than rebuilding the performance from a flat script. It covers 90+ languages. The API version followed on August 6, 2026, with a project-based workflow: an editable JSON transcript and translation, where fixing one line lets you regenerate just that segment instead of re-rendering the whole dub.

Around the same window, ElevenLabs also pushed Scribe v2, a speech-to-text model transcribing 99 languages with word-level timestamps and speaker diarization — which turns out to matter for the next section, because it's also the model everyone got migrated onto.

The v1 sunset: what actually happened July 9

On June 8, 2026, ElevenLabs announced that eleven_monolingual_v1, eleven_multilingual_v1 and scribe_v1 were deprecated, with removal set for July 9. That date came and went as scheduled — anyone still pointing a model_id at one of those three strings after July 9 needed to have already migrated: to eleven_multilingual_v2 for speech, or to scribe_v2 / scribe_v2_realtime for transcription. If you inherited an integration built in 2024 or 2025 and haven't touched the model IDs since, this is the one to check first — it's the kind of break that shows up as a production error in your own logs, not as a warning you'd have seen coming.

Matching the brief on Versely's carried stack

Versely doesn't carry ElevenLabs Music, so if a brief calls for what Music v2 does — a structured, controllable original track rather than a generic loop — the working substitute is Suno Sounds V5.5 (2 credits), which generates with its own looping, tempo and key controls and captures lyric timing as subtitles. It isn't the same chunk-based interface, but the creative target — a track with intentional structure instead of one flat prompt — is reachable with explicit section-by-section prompting the same way Music v2 asks for it structurally. A concrete version of that: instead of prompting with "upbeat corporate background music," prompt section by section — intro: soft piano, 4 bars; build: add light percussion and bass, 8 bars; drop: full arrangement, driving beat, 8 bars; outro: strip back to piano, 4 bars — which gets most of the arrangement control Music v2's chunk plans are designed to guarantee, just expressed as prose instead of a typed API object.

For a hybrid bed — spoken narration sitting on a music-like backing, or a reference-voice-driven read — Seed Audio 1.0 takes a reference audio clip (@Audio1–@Audio3) or a reference image alongside the text, with speed, pitch and volume controls, which covers briefs asking for "this voice, but reading a new script."

For translation and dubbing specifically, the direct read isn't Dubbing v2 at all — it's Versely's own dub-video workflow, which runs a translate-clone-resync pipeline through an ElevenLabs or HeyGen engine (the HeyGen path adds real lip-sync and caps at 8 minutes; the ElevenLabs path skips lip-sync but runs up to 30 minutes with trimming support). A concrete run: take a finished English UGC ad, run it through dub-video with the target language set to Spanish and the ElevenLabs engine, keep the background audio, and you get a re-voiced Spanish cut without re-shooting anything. That's the same job Dubbing v2 does — translate a video without losing the performance — assembled from tools Versely already ships. For the underlying mechanics, see what AI dubbing actually is.

For the TTS layer itself, Eleven Labs Multilingual is already in Versely's catalog at 6 credits per 1,000 characters, running on the current-generation model line — so the July 9 sunset is a non-event for anything built on Versely's own integration; it was never pointed at the deprecated v1 model IDs.

FAQ

Do I need to change anything on Versely because of the ElevenLabs v1 removal? No. Versely's Eleven Labs Multilingual model and dubbing workflow don't run on the deprecated eleven_monolingual_v1 / eleven_multilingual_v1 / scribe_v1 model IDs, so the July 9 removal doesn't touch anything built on Versely.

Is ElevenLabs Music available on Versely? Not currently. For prompt-to-track generation, Versely's catalog runs on Suno Sounds V5.5 instead — accessible through the AI music generator.

What's the practical difference between Dubbing v2 and a plain voiceover swap? A voiceover swap replaces the audio with a fresh read of a translated script. Performance-conditioned dubbing — what Dubbing v2 does, and what the ElevenLabs engine inside Versely's dub-video tool approximates — carries the original delivery's tone and pacing into the new language instead of starting from a flat script.

Why does a transcript-only dub sound worse even when the translation is accurate? Because the translated text carries the words but not the performance. Pauses, emphasis and pacing that made the original delivery land don't exist in a flat script, so a text-to-speech pass has nothing left to reconstruct them from — which is the specific gap performance-conditioned dubbing is built to close.

Takeaway

The headline read on ElevenLabs' summer is "they got faster" — but the more interesting story is "they got more structured." Composition plans instead of a single prompt, performance conditioning instead of transcript-only translation, a hard cutover instead of an indefinite legacy tail. Versely doesn't need to carry every one of these models to hit the same briefs: Suno Sounds V5.5 for original tracks, Seed Audio 1.0 for reference-driven reads, and the dub-video pipeline for translation that keeps the performance intact.