What AI can and can't do with sound

What changed

Until recently, background music for a video meant buying a licence, hours of searching stock libraries, or a studio. Now a track of the right mood and length is made in a minute from a written description.

But the gain isn't where people expect. AI doesn't replace a composer and doesn't write a hit. It covers functional sound — the kind needed to make something work, which nobody will listen to on its own.

What comes out well

Background music for a purpose. Neutral, unobtrusive, the right length and mood.
Short jingles and stings. Three to five seconds of recognisable sound.
Sound effects and ambience. Street noise, footsteps, room tone.
Voicing text. A narrator where there is no narrator.
Drafts and sketches. Testing an arrangement idea before recording with live musicians.

What comes out badly

Matching what you had in mind. A melody in your head isn't reproduced by description.
Repeatability. Getting the same track with a small change is nearly impossible.
Long forms. Over length the structure falls apart and repetition becomes obvious.
Lyrics in languages other than English. Pronunciation and stress drift.
Anything meant as an authorial statement. Functional yes; artistic no.

The practical conclusion

Ask yourself: will anyone listen to this sound separately from the thing it's attached to? If not, generation fits and saves a great deal. If yes, you need a person.

Insight. Separate functional sound from authorial statement. The first generates excellently, the second doesn't, and confusing them is what produces disappointment.

Common mistake. Expecting a melody from your head to be reproduced by description. Words set genre, mood and density, but not a specific melodic line.

Pro tip. The check question: will anyone listen to this separately? A "no" means generation fits and will save you noticeably.

Cheat sheet

The gain is in functional sound.
Good: background, jingles, effects, narration, sketches.
Bad: exact melodies, repeatability, long forms.
The test: is it listened to separately?

1. Where is the gain from sound generation greatest?

2. What do generators handle badly?

3. Which question decides whether generation fits?

Task — checked by AI

List 6 real sound tasks from your practice or plans. Apply the "will anyone listen to this separately" check to each and sort them into: generation fits, needs a person, not sure. For two tasks in the first category, describe what you currently do instead and what it costs in time or money.

🔒 Answer the question correctly to move on to the next lesson.