Eleven v3 vs Eleven Multilingual v2: when each one wins
ElevenLabs v2 vs v3: v3 covers 70+ languages and expressive multi-speaker delivery; Multilingual v2 covers 29 and takes 10,000 characters per request.
Eleven v3 and Eleven Multilingual v2 are two general-purpose ElevenLabs text-to-speech models that look alike in a model picker. They differ on language coverage, character cap and expressive range: v3 covers 70+ languages and audio tags, while v2 covers 29 languages, takes 10,000 characters per request and is the one ElevenLabs calls most stable on long-form generations. Which one wins depends on the script.

AI 音声合成
テキストから自然で表現豊かな音声を生成します。xAI、MiniMax、Inworld、Qwen3(音声クローン対応)など、第一線の TTS モデルから選べます。
The three axes that differ
Language coverage. v3 supports more than 70 languages. v2 supports 29 (ElevenLabs models page, checked 2026-11-10). If your script is in any of the long tail of European, Southeast Asian, or African languages outside the v2 list, the choice is decided before you start: v3 or another vendor.
Character cap per request. v2 accepts 10,000 characters in a single request. v3 caps at 5,000 characters per request on the ElevenLabs API (same page), and the Z.Tools tool capped it at 3,000 while it offered the model. A script that fits in one v2 request needs two v3 requests at the API cap, or four at the 3,000 cap, with the chunking discontinuities that come with that.
Expressive range and audio tags. ElevenLabs describes v3 as built for expressive, dramatic delivery and multi-speaker dialogue, with audio tags that direct emotion or sound effects mid-line. The current models page does not list audio tags for v3, and lists none for v2, so check the tag list for the model you use. The same page describes v2 as most stable on long-form generations and also suited to character voiceovers because of its emotional range. ElevenLabs has reported a 68 percent reduction in errors on complex text for v3 compared with v2; that figure is not on the current models or text-to-speech pages, so it was not re-checked.
When v3 is the answer
The shape of project that wins on v3:
Audiobook scenes with emotional acting. A chapter where a character cries, laughs or shifts register mid-paragraph is the case v3's audio tags are meant for, since they let a writer direct the read at a specific beat. The character cap forces chunking at chapter scale, so cut at scene breaks.
Multi-speaker dialogue with one model. ElevenLabs lists natural multi-speaker dialogue as a v3 use case. An interview, a conversation between two characters or a question-and-answer scene is the fit; this article did not compare speaker shifts between v2 and v3.
Languages outside the v2 list. Polish, Vietnamese, Tamil, Czech, Filipino, Indonesian, Malay and dozens more sit on the v3 supported list and are not on v2. For a script in any of those, the choice is v3.
Content where complex text matters. Long compound sentences, technical vocabulary, dense paragraphs and frequent code-switching are where ElevenLabs's reported 68 percent error reduction (see above) would apply.
Short-form expressive content. Ad reads, character voiceover for animation and podcast intros where the read should land emotionally fit the expressive range v3 is built for. The cap matters less on short scripts.
When Multilingual v2 is the answer
The shape of project that wins on v2:
Long-form single-voice narration. A 90-minute audiobook chapter, a 30-minute course module, a 20-minute corporate explainer, an hour-long meditation guide. The 10,000-character cap means a chapter fits in one or two requests instead of three or four, and ElevenLabs calls v2 the most stable model on long-form generations.
Educational content. ElevenLabs's models page suggests v2 for corporate videos and e-learning materials. For narration where one voice carries the listener through dozens of lessons, request-to-request consistency is the requirement to test.
Brand audio that needs a stable voice. A series of ad reads, a corporate podcast where the same narrator opens every episode, or an onboarding voice that should sound the same months apart needs consistency across regenerations. Whether v2 varies less than v3 here was not measured.
Voice cloning workflows. ElevenLabs has stated that Professional Voice Clones (PVCs) are not yet fully optimized for v3 and recommended Instant Voice Clones or pre-designed voices with v3. That statement is not on the current models or text-to-speech pages, so it was not re-checked. If your project depends on a specific PVC, test it on v3 before committing, and treat v2 as the safer default until you have.
Projects where the cap matters more than the range. Batch jobs where each extra request adds latency, splitting overhead and regeneration risk favor the 10,000-character cap.
The character-cap trap
A 6,000-character chapter (about 1,000 English words, about 6 minutes of audio) generates in one v2 request. The same chapter takes two v3 requests at a 3,000-character limit, or three if the chunking is conservative.
Two requests can produce two slightly different readings of the same voice, with the prosody resetting at the chunk boundary. Breath placement at the join, a register change at the start of the second chunk and pacing when the boundary lands mid-paragraph are where to listen; this article did not measure how audible the join is.
A chunking workflow:
- Split chunks at scene breaks, paragraph breaks, or sentence boundaries, never mid-sentence.
- Keep chunks somewhat under the cap (2,800 characters against 3,000, as an example) to leave room for audio tags in v3.
- Generate adjacent chunks with the same seed when the API exposes one, to keep the voice characteristics consistent.
- Listen across each chunk boundary after generating.
For a long single-voice script, the chunking work is a reason to default to v2 unless v3's emotional range is needed.
Voice cloning and v3
ElevenLabs has stated that Professional Voice Clones (the manually trained, higher-fidelity clones) are not yet fully optimized for v3, and recommended Instant Voice Clones (the faster, lighter clones) or pre-designed library voices with v3. As noted above, that statement was not found on the current ElevenLabs pages. A studio that has invested in a PVC should test it on v3 before moving a production to it, and keep v2 as the default until the test passes.
A new project has no PVC to protect, so it can evaluate both models; if v3's expressive range is the goal, it has no need to start with a PVC.
When the right answer is "use both"
A hypothetical example: a 12-chapter audiobook with two main characters and occasional emotional set pieces. Running v3 for everything costs the chunking work; running v2 for everything gives up v3's expressive range. The split is v2 for the bulk of the narration, where chunks fit in one request, and v3 for the scenes where a character cries, a dialogue lands an emotional beat, or a chapter closes on a moment that needs voice acting. The two are spliced at scene breaks, where content boundaries mask any audio discontinuity.
This is more work than running one model, and this article did not test whether the result is better.

AI 音声合成
テキストから自然で表現豊かな音声を生成します。xAI、MiniMax、Inworld、Qwen3(音声クローン対応)など、第一線の TTS モデルから選べます。

AIカバー曲メーカー
曲をアップロードするだけで、メロディと歌詞はそのままに別のジャンル・声・アレンジのAIカバーを作成。

AI 音声クローン
声をクローンして多言語音声を生成

AI 音声合成
テキストから自然で表現豊かな音声を生成します。xAI、MiniMax、Inworld、Qwen3(音声クローン対応)など、第一線の TTS モデルから選べます。
- ai
- tts
- voice
- ai-tts
- elevenlabs
続けて読む
MiniMax HD vs Turbo vs Eleven Flash for finished work
MiniMax 2.8 HD, MiniMax 2.8 Turbo, and Eleven Flash v2.5 cluster at adjacent per-character prices but split sharply on use case: broadcast finals, fast Chinese agents, and 32-language streaming respectively. Here is which one to pick when.
Dia 1.6B and the case for dialogue-first text-to-speech
Dia 1.6B from Nari Labs is the only dialogue-first text-to-speech model on the AI text-to-speech tool. The architectural difference shows up most in non-verbal cues: real laughter and coughs as audio events, not as read-aloud text. Here is when it wins, and when it does not.
Voice cloning from a few seconds of audio: where it works, where it stops, and consent
Voice cloning from three to ten seconds of audio is now in the AI text-to-speech tool. The technical limits, the legal limits in 2026 (Tennessee ELVIS Act, California AB 2602 and 1836, EU AI Act Article 50), and a consent workflow that holds up.