Multilingual-first publishing in 2026

Ship the language set with the first cut: Seedance lip-sync in 8+ languages, captions in 165 language codes, and a voice layer you can actually buy. Hume Octave 2 is an explainer on this page, not a Versely SKU.

English-first, then a dub sprint, is last year's default. It still wins when the master is a real person, a locked edit, or a signed performance you cannot regenerate. For net-new generated work, the cheaper mistake is waiting. The mouth, the line, and the burned-in type are all language-shaped. If you lock them in English and retrofit, you are paying to fight syllable counts you already chose.

This is not a tour of every localization tool. It is the 2026 publishing order: generate the spoken take in the language it will ship in, caption in that language, and only then decide whether a separate TTS pass is even the job. Octave 2 belongs in the third paragraph of that decision as a speech-language model other teams will name in briefs. Versely does not serve it.

Generate the line in the language

Seedance 2.0 is the video row this page is actually about. ByteDance's model takes unified multimodal input — text, image, video, audio, up to 12 references — and generates up to 15 seconds with phoneme-level lip-sync in 8+ languages (fal's Seedance 2.0 card; Atlas Cloud's 2026 cost write-up; Morphic on Seedance 4K). Native 480p/720p is what the technical report describes; 4K shows up on some hosts. On Versely, quote the credit row, never a host's dollars.

Phoneme-level lip-sync is the publishing fact. The mouth is not a later pass glued onto an English performance. The model is producing the visemes and the audio for that language in the same generation. Spanish, Japanese, Hindi — whatever set Seedance actually covers in the 8+ — get a take whose timing belongs to that language. German running long, Japanese compressing, English sitting in the middle: those are not dub artifacts you stretch. They are the shot.

That is why "we'll translate the VO later" is the wrong first sentence for generated talking shots. Later means you already spent the motion pass on English visemes. A dedicated dub-and-lipsync pipeline can still reshape a mouth. It is a harder problem than generating the mouth for the language you meant. Use the retrofit when the face is a real person you cannot resample. Use Seedance (or another audio-native talking row) when the face is generated and the language is a first-class input.

Native audio is the category this sits in. Seedance 2.0, Veo 3.1, and Kling 3.0 all produce synced sound in one pass; Veo is the one trackers still single out for 48 kHz dialogue (Magic Hour; wavespeed). If the job is "she says this line, in this language, on camera," write the line in that language in the prompt. Do not write the English line and hope a caption translates the performance.

References travel across languages better than adjectives do. The same three product stills, the same wardrobe plate, the same room — those 12-reference slots are how a Spanish take and a French take remain the same SKU. The prompt changes. The pack does not. That is multilingual-first as a stills discipline, not as a slogan.

Duration is a constraint, not a vibe. Fifteen seconds is the Seedance 2.0 ceiling on the catalog row. A two-minute explainer is not one generation in five languages. It is a cut list: a talking open, a product pass, a captioned recap. Each beat can be native in-language. The assembly is still an edit.

Captions are a second language layer

Spoken language and readable language are not the same deliverable. A Seedance take in Portuguese still needs type for silent autoplay, for accessibility, and for the feed that never turns the sound on. Captions are not "the translation." They are the on-screen language of that file.

Versely's VEED captioning pass transcribes against 165 language codes — en-US, es-MX, ja-JP, ar-SA, and the rest of that table. It transcribes what is spoken. It does not translate a Spanish take into English type. If you need a different language on screen than in the mouth, that is a translation job, not a transcription default. The AI caption generator is the landing page for the burn-in path. Add captions to a video is the editing task. Caption styles is the look, which should come from the brand kit rather than from whichever preset the last editor liked.

Multilingual-first means you pick the caption language when you pick the take, not at upload. A Japanese talking shot with English captions is a different product from a Japanese talking shot with Japanese captions. Both are valid. Mixing them by accident is how a market feels like an afterthought. Burn the type into the file you will actually post. Sidecar captions are a different delivery; social feeds still eat the burned-in frame.

Safe area and line length do not become optional because the language changed. CJK, Arabic, and long German compounds all blow a layout that was designed around short English hooks. If the kit has a caption style, use it. If the kit assumes English line breaks, test one file per language before you batch. The 165-code table is a transcription capability. It is not a typesetting guarantee.

Do not caption the English master and then stretch those timings onto a dubbed mouth. The timings belong to the audio you are posting. A phoneme-level Seedance take already has different peaks than the English cousin. Transcribe that file.

Octave 2 is an explainer, not a SKU

Hume launched Octave 2 as a multilingual speech-language model: 11+ languages (English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi, Arabic), 40% faster than Octave 1 with generation under 200 ms, 50% cheaper than Octave 1, multi-speaker, voice conversion, and phoneme editing (Hume's Octave 2 launch post; TestingCatalog's Octave 2 note). Those are Hume's claims, dated to that launch. They are the reason briefs will say "Octave 2."

Versely does not serve Octave 2. There is no /models/octave-2, and Hume is sunsetting its TTS and EVI APIs: access ends on 13 November 2026 (Hume changelog). A project built around Hume's phoneme editor or that specific 11-language set has to run on Hume's API until then, and should not start a new multilingual stack on it. Hume's description field for plain-English acting directions is documented for Octave 1 only, with Octave 2 support listed as "coming soon" (Hume docs). This site will not route it, will not bill it in credits, and will not pretend a clone is the same object.

What Versely does serve, for the voice layer of a multilingual stack:

Cartesia Sonic 3.5 — the low-latency read, the volume default when the register is informational
Inworld TTS 2 — character and game-adjacent reads, when the voice is part of the scene
Gemini 3.1 Flash TTS — Google's Flash TTS row: multilingual synthesis, multi-speaker, inline audio tags

ElevenLabs is the other name briefs will use. If it is not a published Versely row you have grepped, do not invent a model page for it. Compare Octave 2 against the four voices this studio actually meters, not against a wish list.

The publishing rule is crude:

The face is generated and the line is on camera. Seedance (or Veo) in that language. Captions in that language. TTS is not the hero.
The face is real, or the picture is already locked. Dub and lipsync, or a voiceover laid under b-roll. This is where Sonic 3.5, Inworld TTS 2, or Gemini 3.1 Flash TTS earn their keep. Octave 2 sat in this bucket for teams on Hume, until its API closes on 13 November 2026.
The face is generated but you need a specific cloned voice. Silent or mute picture, then a dedicated TTS pass, then attach. Native audio will not be your clone.

Phoneme editing, on Octave 2, is a TTS control: fix a name, a brand, a number, without rerunning the whole read. Seedance's phoneme-level lip-sync is a video control: the mouth agrees with the language. Do not mash those two "phoneme" sentences into one capability. They are different products.

First cut, every market

Multilingual-first is a calendar, not a translation ticket at the end.

Write the brief once, with language as a column: which markets ship this week, which language the mouth uses, which language the captions use, which stills are shared. Generate the talking takes per language on Seedance while the references are still the signed ones. Caption each file in its language. Only then cut the English-only version, if English is even in the set.

The cost of that order is more generates. The cost of the other order is a master that does not fit the second language, a dub that fights the cut, and captions timed to the wrong peaks. Teams already report up to 11× more video per month with AI workflows (OpusClip; Hedra; Pictory). Volume is not the scarce resource. A language-shaped first cut is.

Do not ship five languages of the same hook if you have not watched one file per language. A viseme error, a caption collision, a brand name pronounced as English inside a Spanish line — those are first-cut bugs. They are cheaper than a dubbed master with a year of English muscle memory.

When the next clip should exist in more than one language on the first pass, sign up, lock the references, and generate the mouth in the language you are actually posting.

FAQ

Does Versely sell Hume Octave 2?

No. Octave 2 is Hume's multilingual TTS, and Hume is closing that API on 13 November 2026. This page explains it because briefs name it. For voice on Versely, use Cartesia Sonic 3.5, Inworld TTS 2, or Gemini 3.1 Flash TTS. For talking video, use Seedance 2.0 (or Veo) and write the line in the target language.

Are 165 caption languages the same as Seedance's 8+ lip-sync languages?

No. 165 is the transcription code table on the caption pass. 8+ is Seedance's phoneme-level lip-sync set. You can caption a file in a language the mouth was never generated in. That is a different product. Match them on purpose.

Should we still dub a real founder into other languages?

Yes. A real person is not a Seedance take. Dubbing and lipsync exist for locked, human performances. Generate-native in-language is for net-new generated talking shots, where the mouth can be born in that language.