# MiniMax H3 dialogue workflow: brief, audio and sync review

A video with an audio track is not necessarily a usable dialogue clip. This tutorial separates the script, scene direction and audio review for `minimax-h3`. It is a production checklist, not a claim about measured speech accuracy or lip synchronization.

## Write a brief that fits one short scene

Use one speaker and one short line for the initial sample. Describe the visible action, intended language, exact spoken words and background sound separately. Long scripts and multiple overlapping actions make diagnosis harder: a missing word may be a timing problem, a generation error or an unrealistic brief.

| Brief component | Example | What to verify |
| --- | --- | --- |
| Visual action | A presenter places a mug on a table | Object and action match the brief |
| Spoken line | “Welcome to the studio.” | Correct words, no extra sentence |
| Language | English | Intended language and intelligibility |
| Background | Quiet room, no music | Speech remains audible |
| Timing | One continuous short scene | No clipped opening or ending |

These are acceptance goals, not guaranteed output characteristics. If exact narration is mandatory, consider an independently reviewed voiceover and editing workflow; do not assume repeated generation is the only correction method.

## Use platform fields, not an invented audio switch

The repository H3 adapter documents native audio behavior, and its schema has no `generate_audio` switch. Describe audio intent in the prompt; do not add an undeclared toggle. The following task body uses only supported fields. Confirm the live schema and access before a paid call.

```json
{
  "model": "minimax-h3",
  "input": {
    "prompt": "One continuous shot in a quiet studio. One adult presenter gently places a plain mug on the table and says in English exactly: Welcome to the studio. Clear natural speech, quiet room ambience, no music, no other voices and no additional dialogue. Keep the full action visible.",
    "seconds": 8,
    "resolution": "768P",
    "aspect_ratio": "16:9"
  }
}
```

Submit to `POST /v1/tasks` with the authentication in the [H3 guide](/docs/models/minimax-h3), save the ID and poll that task. The eight-second duration is an example brief choice, not a speech-speed guarantee. This request contains no audio reference and does not request voice cloning.

If you later use authorized audio references, inspect the exact H3 input-mode combination rules first. `audios` is a reference field, not a separate transcription or dubbing endpoint. H3 Max Turbo's platform schema does not declare the same general audio-reference input; see [H3 / Turbo selection](/docs/media/minimax-h3-selection).

## Listen and watch as separate review passes

First verify that the returned asset opens and contains the expected audio track. Listen without watching: check the exact words, language, clipped syllables, unwanted voices, music and audible distortion. Automated transcription can help flag discrepancies but cannot replace listening, especially for names or short utterances.

Then watch with sound. Check whether speech timing makes sense against the visible action and mouth movement, and whether the end of the line is cut off. Inspect the clip in the actual destination player; playback behavior can differ after export. Do not infer good synchronization merely because audio and video have the same duration.

Record audio correctness, visual correctness and synchronization independently. A technically completed task may fail one or more of these criteria. Keep original files before editing, and validate the final export again after cuts, fades or voiceover replacement.

## Budget corrections explicitly

Classify each issue: missing/invalid asset, wrong spoken words, visual drift, or timing mismatch. Correct the smallest part of the workflow that can meet the brief. If you generate again, record a new attempt and charge; do not resubmit automatically after a polling deadline. Use the [task workflow](/docs/media/video-nodejs), [shot planning guide](/docs/media/minimax-h3-shot-continuity) and [budget guide](/docs/media/batch-budget).

Source and contract check: 2026-11-03, [MiniMax native video API](https://platform.minimaxi.com/docs/api-reference/video-generation-v2-create), platform H3 schema and adapter. Native concepts do not imply every upstream parameter is exposed. No paid speech or video test was performed for this article.