MiniMax H3 Native Audio: How to Get Dialogue and Sound Right

The defining break between MiniMax H3 and the earlier Hailuo line is sound. Hailuo 2.3 outputs silent footage that needs a separate audio pass; H3 generates stereo dialogue, ambience and music in the same generation as the picture. Because sound and frames come from one pass, timing cues — speech, impacts, footsteps — stay aligned. Here is how to get the most out of it.

Write audio into the shot description

The single biggest lever: describe sound explicitly instead of leaving it implicit. A reliable structure is visual action + quoted dialogue + ambience + delivery note:

A mechanic wipes her hands and says "she will run another hundred thousand miles," dry and confident; garage reverb, a radio playing faintly.

• Dialogue: put lines in quotes; short sentences sync better than long monologues.
• Ambience: name the texture — leather interior, outdoor wind, cafe murmur.
• Silence: say so when it is intentional ("no dialogue, only waves").

Voice transfer from a reference clip

H3 accepts up to 3 reference audio clips in an omni-reference request. Feed it a clean voice recording and the generated character speaks in that voice. This is ideal for consistent brand narrators across a campaign. Reference audio is billed free, so this control costs nothing extra.

Music and score

You can request a musical bed ("soft jazz," "ambient synth pad") and H3 will compose one that fits the cut. For publish-grade music you may still prefer a licensed track, but for drafts, shorts and internal review the in-pass score is remarkably usable.

[INSERT_IMAGE: screenshot of an H3 generation result with the audio waveform visible, annotated to show the dialogue line matching the lip movement frames]

Common problems and fixes

• Words mumble or drift: shorten the line and add a delivery note ("calm," "urgent whisper").
• Ambience overwhelms dialogue: explicitly rank them ("dialogue forward, faint street noise").
• Accent or pronunciation off: use voice transfer from a reference clip instead of describing the voice.

Always give generated audio a human review pass before publishing — wording, pronunciation and timing artifacts still slip through.

How it compares

Native audio is becoming table stakes — Veo 3.1 is the benchmark for lip-sync realism, while H3 counters with 2K output and a much lower per-second price. See the head-to-head in MiniMax H3 vs Veo 3.1, and read the full H3 review.