H3 Max Custom Audio: Target, Reference, or Lip Sync?

Yes, H3 Max can work with custom audio—but “use this audio” can mean four different jobs. The standard video routes can generate a new soundtrack from the prompt. The current Turbo API also has a target-audio input. A separate reference route can use an audio clip as creative guidance, while the lip-sync route drives a face from a supplied recording.

Those inputs are not interchangeable. Choosing the wrong one is how a voice reference turns into an unwanted replacement track, or an exact recording becomes only a loose influence.

This site's generator currently supports the first job: it creates original dialogue, ambience, effects, and music from written direction. It does not show an audio-upload control or send a target-audio, reference-audio, or lip-sync request.

Pick the audio job before you pick the endpoint

The practical difference is replacement versus influence. If a client has approved one finished voiceover and that exact file must play, reference audio is the wrong contract. If you want qualities such as pace or atmosphere to guide a newly generated scene, target audio may be too literal.

Target audio pins a supplied soundtrack to the generated clip

The current H3 Max Turbo text-to-video schema and image-to-video schema expose target_audio_url. In the live documentation checked on October 12, the audio must be at least two seconds and no larger than 15 MB.

The timing behavior is worth planning around:

• if the audio is longer than the requested video, the end of the track is trimmed;
• if it is shorter, the remaining video is padded with silence;
• the audio is not sped up or slowed down to fit;
• it replaces the soundtrack H3 Max would otherwise generate.

That makes target audio useful for a cleared voiceover, an original performance, or a finished sound bed whose beginning must stay intact. It does not provide editable speech, music, and effect stems. If you need to move one word, lower music under dialogue, or preserve a precise ending, assemble the final mix in an editor.

Reference audio is a direction, not a locked soundtrack

The H3 Max reference-to-video schema accepts audio references from two to 15 seconds each, with no more than 15 seconds of reference audio in total. Image, video, and audio references share a maximum of 12 files. The prompt can address them by order—Audio 1, Audio 2, and so on—just as it addresses image and video references.

There is a documentation wrinkle. On October 12, fal's live schema described image, video, and audio as usable individually or together, while its current H3 Max workflow tutorial said an audio reference must be paired with at least one image or video. Until those first-party descriptions agree, the safer portable workflow is to pair authorized audio with an authorized visual reference and check the live schema before building an integration.

Write one role for every file. For example:

Image 1 defines the adult radio repairer, workbench, clothing, and warm afternoon light. Audio 1 guides the repairer's calm speaking cadence and the soft room tone. Create a new five-second medium close-up. The repairer looks up from the open radio, says one short original line, then turns a tuning dial. A single clear chime follows the hand movement. Keep the camera still and preserve the quiet, intimate pace.

This is a planning template, not a tested output claim. The useful part is the assignment: ask the image to guide appearance and setting, ask the audio to guide cadence and atmosphere, and use the prompt to define the new action. Do not ask one reference to be identity, dialogue, music, camera movement, and exact timing all at once.

Lip sync is a separate talking-face job

If the real requirement is “make this face speak this recording,” use the dedicated H3 Max lip-sync schema, not reference audio. That route requires an image and an audio URL. The current schema sets a five-second minimum for the audio, and the output duration follows the supplied track; check the live endpoint for its current upper limit.

Lip sync solves a narrower problem than general video generation. It does not turn one recording into a polished scene with cutaways, a designed mix, and several speakers. Start from a face that is large enough to read, unobstructed, and close to front or three-quarter view. Keep the first test to one speaker. The lip-sync troubleshooting guide shows how to separate wrong words from bad mouth timing before paying for another attempt.

For original sound, direct the moment instead of naming a mood

You do not need an uploaded track to make a useful sound brief. Tie each sound to a visible event, give speech its own beat, and remove sounds that would compete with the focal moment.

For the radio-repair scene, a prompt-only version could read:

Five-second medium close-up of an adult radio repairer at a quiet workbench in warm afternoon light. The camera remains still. After one second, the repairer looks up and says once, in a calm natural voice: "I found the signal." The lips close, the right hand turns the tuning dial, and one soft electronic chime sounds exactly with the turn. Audio: close clear speech, low room tone, one chime after the sentence. No music, narration, other voices, or cut.

The line, hand movement, and chime create three review points. If the result fails, you can say whether the wording, the visible cue, or the sound timing missed. “Cinematic audio” gives you nothing that specific to diagnose. The synchronized-audio guide has more patterns for dialogue, ambience, effects, and music without overloading one short clip.

A five-question check before you spend on another take

1. Must the final file contain this exact recording? Choose target audio or finish in an editor; do not rely on a reference.
2. Do you only want the voice, rhythm, or atmosphere to guide a new result? Use reference audio and assign its role in the prompt.
3. Must a visible mouth follow the recording? Use the lip-sync route with a suitable portrait.
4. Would original sound be acceptable? Use prompt-generated audio and make each sound observable.
5. Do you need stems or frame-accurate mixing? Generate the picture, then move to an editor.

Use only audio and images you have permission to upload. A publicly available song, performance, or voice is not automatically licensed for a new commercial clip. The endpoint determines how an asset is processed; it does not grant the rights to use it.

For this site, the boundary is simple: open the standard generator when original synchronized sound fits the brief. If an exact recording, uploaded voice reference, or dedicated lip-sync pass is non-negotiable, plan that separate workflow before generating the picture.