The first take I rendered had rain in it. I had not asked for rain. I had asked for a wet street at night, and the model read that as sound design rather than colour grading.
My original plan was the usual one: generate silent, then dub. H3 does not really do silent. Every render carries an audio track, whether or not the prompt mentions sound, and once you stop fighting that, how you write a scene changes.
The short version. Every H3 render carries an audio track, reference audio is free up to three clips of 2–15 s, and the output audio arrives at 32 kHz.
What can go in the three audio slots?
Up to three audio references can go in one request, each two to fifteen seconds, mp3 or wav (flac on the standard tiers), fifteen megabytes each. They are addressed in the prompt the same way images are: Audio 1, Audio 2, Audio 3.
Reference audio is not billed. The per-second rate covers the output.
I use the slots for three different jobs now, and I try not to mix them:
- Voice. One line read the way I want it performed. The reference carries register and pace; the prompt carries the words.
- Ambience. A room tone or a street bed. Two seconds of the right room is worth more than a paragraph describing the room.
- Rhythm. A four-bar loop, used carefully. It is not a licence to demand beat-synced cutting — output durations land on a 24 fps frame grid, so a five-second request renders about 5.167 seconds. Cut to the timeline, not the other way round.
Writing sound like camera direction
Sound responds to the same discipline as camera notes: name the source, keep it short, do not stack.
Rain on a metal awning. A door closes off-screen. She says, "You're late," quietly, not looking up.
Two rules came out of my own failures. Put the line before the reaction — otherwise the model animates the pause and then invents a second line to fill it. And never describe more than two simultaneous sources; it is mixing, not following a cue sheet.
Why is the output audio only 32 kHz?
The audio arrives at 32 kHz. Most delivery specs assume 44.1 or 48. Platforms will re-encode it, editors will accept it silently and sound slightly dull, and neither will tell you why.
A deliberate resample before delivery takes a second and removes the variable. It is the least glamorous step in the whole pipeline and the one I now never skip.
When to strip the track
If the shot is a silent product turn, an ambient bed is not free — the model will add motion to justify the sound you asked for. Strip it, but strip it on purpose.
If the cut is already locked to music, generate the picture and lay the track in the edit. The model will otherwise try to sync to a reference it only partly understands, and you get edits that feel almost right.
And if the shot needs first and last frame conditioning, audio references are not available at all — that mode cannot carry them. Those are separate routes, and mixing them is rejected.
Audio references, tiers and rates are on the H3 page; the audio studio has its own page.
I work on Callirra, which is how I can see this behaviour rather than infer it, and why the takes I learn from are the ones on our own page. If you have measured the output sample rate on another H3 lane and got a different number, that is worth posting — a rate table from three lanes beats mine from one.