AI music for TikTok: describe it, and a track comes back

The input is a description in your own words. Dark trap with heavy 808s and a detuned piano is a valid brief. So is warm acoustic guitar, gentle, no drums.

Suggestion chips write into that box rather than replacing it, so you can start from upbeat pop and then add what you actually mean.

The lyrics box is optional. Empty means instrumental. Anything in it gets sung.

Pick a length from 15, 30, 60, 90 or 120 seconds. The price does not change with length: 20 credits either way.

The 300 free credits with no card cover fifteen tracks, which is enough to find a sound rather than to accept the first one.

Basic is $39.99 a month for 2,500 credits, which would be 125 tracks if you spent a whole plan on music, and nobody does.

The track is generated for you rather than licensed from a catalogue, which is a different legal shape from buying a stock licence.

Nothing was matched, cleared or sublicensed, and no other advertiser has the same recording sitting in their library.

That removes the common failure where an audio matching system flags a track you paid for, because the recording did not exist before you asked for it.

It does not amount to a permission slip, and any page telling you otherwise is spending trust it has not earned.

Confirm the usage terms that apply to your own account before a track runs behind paid spend. Keep the description, the date and the file.

That record takes ten seconds to save and is the only thing that answers a question asked eighteen months later.

A phone speaker is a few millimetres across. It reproduces almost nothing below roughly 500 Hz, and most viewers never plug in headphones.

So the 808 that shook your monitor is inaudible on the device the ad is actually watched on.

What survives is the midrange, which is exactly where a human voice lives, and that is where the problem starts.

A track described as huge and full arrives packed with midrange. It competes directly with whatever anybody is saying over it.

Ask for space instead. Sparse, mostly low end and high percussion, nothing in the middle is a better brief than anything with the word big in it.

Then listen to the finished ad on a phone at arm's length before you post it, which is how the audience will meet it.

A sung vocal and a spoken voice occupy the same frequency range and the same part of a listener's attention.

Run both at once and the viewer hears neither. The brain cannot parse two sets of words at the same time.

So use lyrics where nobody is talking, and use instrumental everywhere a voice-over runs.

The one exception that works is a sung line in the first two seconds before the talking starts, then instrumental underneath.

That means two tracks for one ad, at 20 credits each, which is a cheaper fix than any amount of level adjustment.

In the editor the music controls are turn it on, pick the track, and set how loud it sits. There is no ducking and no stem separation to rescue a bad pairing.

Sixty credits buys three tries at a sound you have never made. One attempt covers a sound you have already found and are repeating.

That is the working budget, and it is small enough that a wrong answer costs nothing worth arguing about.

Four plain limits sit around it. A returned track cannot be edited, retimed or stripped into stems.

If it is close but wrong, change the description and generate again for 20 credits. That is how the module expects to be used.

The longest track is 120 seconds, so anything longer has to be assembled elsewhere.

The price does not scale down either. Fifteen seconds costs the same 20 credits as two minutes.

And a specific recognisable song is not on offer. That is a licence negotiation, and it is expensive for a reason.

Platform catalogue music has the same problem in reverse. That permission is usually scoped to organic posting and does not stretch to paid advertising.

One more thing worth checking before you spend anything. If your ad is a talking head with no gaps, play it with the track muted first.