MiniMax H3's Native Audio: One-Take Sound for Product Demos

Siray.AI

08 Sep 2026 • 4 min read

MiniMax H3's Native Audio: One-Take Sound for Product Demos

Key Takeaways

MiniMax H3 (minimax/minimax-h3-t2v, -i2v, -ref2v) generates its own matching audio track alongside the video in the same call — it does not accept a voiceover script or audio file for the video to sync to.
Checked against docs.siray.ai's live OpenAPI specs, prose docs, and llms.txt on 2026-09-06: zero mentions of "sync," "lip-sync," "voiceover," or "dubbing" across all three endpoints.
ref2v's optional audios field takes up to 9 reference audio URLs alongside images/videos — Siray's own documentation describes it only as "array of input audio URLs," with no stated lip-sync function.

Most product demo videos still get built the slow way: someone writes a script, records a voiceover, drops the clip into an editor, and manually drags the audio track until it lines up with the footage. That alignment pass is its own line item — and it's the part MiniMax H3 actually removes, not the part its name might suggest.

What H3 Actually Does With Audio

Here's the mechanism, checked directly against Siray's live API documentation for all three H3 endpoints (minimax/minimax-h3-t2v, minimax/minimax-h3-i2v, minimax/minimax-h3-ref2v) on 2026-09-06: none of the request schemas, prose model pages, or the llms.txt index mention "sync," "lip-sync," "voiceover," or "dubbing." Nowhere in the documentation does H3 accept an external voiceover script or audio track and align a generated video to it.

What it does instead: when H3 generates a video from a prompt, it produces its own matching audio track — native stereo sound, generated alongside the visuals in the same call. There's no separate step where you feed in a script and get back lip-synced footage. The audio is a product of the same generation pass as the video, not a post-processing sync operation.

The one field worth clarifying is ref2v's audios parameter — an optional array accepting up to 9 audio URLs, sitting next to images and videos as reference inputs. Siray's model page describes ref2v's reference guidance only as something that "steers subjects, style, and scenes" — visual language, even though the schema also lists audios. The field's own description in the spec is a bare "array of input audio URLs," with no further detail. Treat it as an optional style reference, not a script you hand over for the model to lip-sync against — the documentation doesn't support that use.

Setting Up a Demo Prompt, Element by Element

Pick the right endpoint

minimax/minimax-h3-t2v — pure prompt-to-video. Good when you're describing a scene from scratch (a product concept, a UI you haven't screenshotted yet).
minimax/minimax-h3-i2v — animates a static image. Required fields: model, prompt, image, duration, size. This is the natural fit for demos: feed it an actual app screenshot or product shot and describe the motion and narration you want layered on top.
minimax/minimax-h3-ref2v — adds optional images/videos/audios reference arrays (up to 9 URLs each) for extra style guidance.

Mini example: a two-panel SaaS dashboard screenshot into i2v turns a static UI into a moving walkthrough, without hand-animating each panel.

Write the sound into the prompt text

There's no dedicated script or voiceover field on any H3 endpoint. Tone, mood, and sound cues only reach the model through the prompt itself.

Mini example prompt: "A clean SaaS dashboard demo, calm confident narration describing a one-click export feature, soft ambient background music, no overlapping dialogue."

Set size and duration for a demo cut

The live size enum on all three endpoints, as of this check, is 768p and 2k — 2k for a landing-page hero clip, 768p for a quick social teaser. duration runs 5 to 15 whole seconds. That's enough for one feature callout per call, not a full multi-feature reel in a single generation. Note: pricing pages elsewhere reference a 4K tier for H3, but it isn't in the current selectable size enum — don't plan a demo cut around a resolution the API doesn't currently expose.

Use the ref2v audio reference sparingly

If reaching for ref2v's audios array, treat each URL as one more style signal alongside images/videos — not a voiceover track the output will match word for word. Siray's documentation doesn't specify precision here, so don't promise stakeholders frame-level control.

Full example

Endpoint: minimax/minimax-h3-i2v
image: product screenshot URL
prompt: "Smooth zoom into the dashboard's export button, upbeat narration explaining the one-click export feature, light corporate background music"
size: "2k"
duration: 10

Where This Fits

Feature-highlight clips for release notes
Product walkthrough videos for onboarding pages
Landing-page hero videos
Short launch teasers for social

Developer Notes

All three H3 endpoints post to the same /v1/video/generations path and return the same async task pattern — one Siray API key covers t2v, i2v, and ref2v by changing the model field, not the integration code. That's the practical version of "one API key, one line change": swapping endpoints for a different demo format doesn't mean rebuilding the request.

Summary

The time H3 saves isn't the narration itself — it's the separate pass of recording sound and lining it up with footage afterward. Picture and matching audio come out of one call; what goes in is still just a well-written prompt.

Create your free Siray account and generate a product demo with MiniMax H3's native audio in one call.