What is Native audio (audio-video generation)? Definition, key points and related tools
The video model generates speech and ambience together with the picture, lip-synced to whoever speaks on screen. Supported by MiniMax H3, Seedance 2.x, Wan 3.0, Veo 3.1, Kling 3.0 and Vidu Q3, among others (Sora 2 also, until its API shuts down on 2026-09-24); models without it need post-production voice and lip-sync.
What changes when the video model does the talking
With native audio, the spoken line, the mouth shapes, room tone and effects come out of one generation. That removes the separate lip-sync pass that text-to-speech laid over a finished picture needs, and it also welds the sound to the picture: a mispronounced word or a dragged syllable means regenerating the whole segment at the video rate.
Vendors describe the sound side differently. MiniMax describes H3's output as native stereo in which voice, sound effects and music are jointly modeled with no separation between them (MiniMax, vendor claim); ByteDance describes Seedance 2.0 as two-channel stereo with background music, ambient effects and character voices output as parallel tracks (ByteDance Seed, vendor claim). If you plan to change the score or re-voice one line in the edit, find out first whether the product you generate in hands you those parts as separate stems or as one mixed track. For the choice between preset voices, cloning and native audio as whole routes, see the AI voice explainer.
How to judge a model's native audio
| Model | Length per call | Audio switch and price | Dialogue notes |
|---|---|---|---|
| Veo 3.1 | 4, 6 or 8 s; 1080p and 4K at 8 s | Audio included in the per-second price; blocked videos not charged | Quote the spoken line (Gemini API) |
| Wan 3.0 | 2–30 s | On by default; switching it off does not change the price (Alibaba Cloud) | "No dialogue." and "No background music." as off switches (prompt guide) |
| MiniMax H3 | 4–15 s | No separate audio line item published; reference audio input free | Up to 3 reference voices (API docs) |
| Seedance 2.0 | 15 s multi-shot output (vendor claim) | Not disclosed in the launch post | Vendor cites better handling of dialects and singing; language list not disclosed (ByteDance Seed) |
| Kling VIDEO 3.0 | 3–15 s | Not disclosed | Chinese, English, Japanese, Korean, Spanish (Kling blog) |
| Vidu (API) | Per model | Audio option adds 15 credits per image- or reference-to-video task (Vidu) | Not disclosed |
| Sora 2 (API) | Up to 20 s per generation | Output includes audio (OpenAI) | Not disclosed |
Beyond the table, five questions decide whether a model suits dialogue-heavy drama: does it put the line in the right mouth in a two-shot; are your languages documented; can you silence dialogue and music explicitly; does audio change the price; and does one call fit your longest line plus a reaction beat?
Writing lines the model will perform
- Quote the line and name the speaker. Google's Veo guidance is to use quotation marks for specific speech. Alibaba's Wan formula builds a voice from the character's line plus emotion, tone, speed, timbre and accent.
- Anchor speech to a visible action. For clips with several speakers, Wan's guide recommends a unique label for each character instead of pronouns, describing the action before the speech, distinct voice descriptions per character and explicit sequencing words.
- Give every line its seconds. Spare time with nothing scheduled gets filled by the model; see dialogue timing.
- Write silence as an instruction. A shot with no speech should say so where the model has a control phrase for it.
Shot-level wording in general is in the prompt-writing explainer.
Common mistakes and why they happen
- A line repeated or stretched. The segment is longer than the dialogue and no action fills the gap, so the model fills it with speech. Write a reaction or movement into the surplus.
- The wrong mouth moves. The speaker was not named, or two faces are equally prominent. Name the speaker first and keep one talking face per shot.
- Music nobody asked for. Some models add a score unless the prompt rules it out. Ask for no background music and add score in the edit.
- A voice that changes between segments. Each call invents timbre unless it is given one; bind a reference audio clip per character.
- TTS pickups on a native-audio character. A patched line in a different synthetic voice is heard at once. Re-generate, or cut to a reaction shot and keep the pickup off-screen.
- A blocked render read as a bug. Google notes Veo 3.1 sometimes blocks a video because of safety filters or audio processing issues, and does not charge for it (Gemini API); see content moderation.
Pipeline tools inherit these behaviors from the model tier they call. SceneMixer, for example, says it matches each character to a preset voice and attaches it as reference audio (SceneMixer guide, vendor claim), so the same checks apply to whichever tier renders the segment.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| Veo 3.1 price and audio | Standard $0.40/s at 720p or 1080p, $0.60/s at 4K; audio included; not charged when a video is blocked | ai.google.dev | 2026-09-12 |
| Wan 3.0 audio parameter | Audio on by default; turning it off does not affect pricing; 2–30 s per call | www.alibabacloud.com | 2026-09-12 |
| Vidu API audio option | +15 credits per image- or reference-to-video task (1 credit = $0.005) | platform.vidu.com | 2026-09-12 |
| Kling VIDEO 3.0 native audio | Chinese, English, Japanese, Korean and Spanish; 3–15 s per generation | klingai.com | 2026-09-12 |
| MiniMax H3 audio (vendor claim) | Native stereo; voice, effects and music jointly modeled, not separated | www.minimax.io | 2026-09-12 |
FAQ
Which video models generate speech natively?
Google documents native audio for Veo 3.1, Alibaba for Wan 2.5–3.0, MiniMax for H3, ByteDance for Seedance 2.0, and Kling for its VIDEO 3.0 series; OpenAI's Sora 2 output includes audio. Vidu's API lists an audio option priced at 15 extra credits per task. Older versions on the same platforms can be silent, so check the model version, not the brand.
Is native audio more expensive than adding TTS afterwards?
Per character, TTS is cheaper: MiniMax lists HD speech at $100 per million characters. Native audio often has no line item (Veo 3.1 includes it; Wan 3.0 prices audio on and off the same), but every retake of a line is paid again as video seconds, so its real cost is your regeneration rate.
Can I fix one word or swap the music afterwards?
Not inside the generated track when voice, effects and music are mixed together, as MiniMax describes for H3. Plan for it: ask for no background music at generation and score in the edit; fix a single word by regenerating the segment, or by cutting to a listener and placing the corrected line off-screen.
Sources
- https://ai.google.dev/gemini-api/docs/veo
- https://ai.google.dev/gemini-api/docs/pricing
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://www.alibabacloud.com/help/en/model-studio/text-to-video-prompt
- https://www.minimax.io/blog/minimax-h3
- https://platform.minimax.io/docs/api-reference/video-generation-v2-create
- https://platform.minimax.io/docs/guides/pricing-paygo
- https://klingai.com/blog/kling-video-3-omni-native-lip-sync-audio-guide
- https://platform.vidu.com/docs/pricing
- https://seed.bytedance.com/en/blog/official-launch-of-seedance-2-0
- https://developers.openai.com/api/docs/guides/video-generation
- https://scenemixer.com/guides/wan-3-short-drama/