Glossary
What is AI music generation?
In short
AI music generation is the creation of melodies, arrangements and songs by a neural network from a text description, a hummed tune or a style sample.
Images, video, audio · AI music generation
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine an orchestra of soft toy animals: a bunny on the drums, a bear on the trumpet, a cat at the piano. You tell them, “Play something happy about the sea”, and together they make up a new tune. A computer program can do much the same: from a few words, it writes music nobody has heard before. And real musicians teach it and check whether the result sounds good.
In plain words
The neural network has listened to a huge amount of music and learned how rhythm, harmony, the sound of instruments and the voice work. You write a request, such as “calm jazz for a cafe, piano and double bass, no vocals”, and the model builds the sound step by step. Some systems give a finished recording at once, others write notes first. Many can continue a hummed tune or change the style of a finished piece. Give some services the lyrics, and they will generate the vocals too. The result sounds like studio work, but listen carefully: there are strange transitions and unclear words.
For business
Companies use generated music as background for videos, podcasts, ads, phone menus and shops, where they used to buy licenses for stock tracks. The price is usually the service subscription plus a plan that allows commercial use: on free plans, rights are often limited. The main risks are legal. Service terms differ: who owns the track, whether you can register and monetize it, and what happens if it sounds like someone else’s song. The model may repeat a recognizable motif by chance, and a voice that sounds like a famous singer creates a separate dispute. Before launch, a lawyer reads the terms of the chosen service, and marketing keeps the prompts and generation dates in case of claims.
For a specialist
Music generation is solved in two ways: symbolically, when the model produces notes or MIDI, and in audio, when it produces sound. Modern audio systems compress the signal with a neural codec into discrete tokens and model them with a transformer. MusicLM (Agostinelli et al., 2023) treats the task as hierarchical sequence modeling, outputs 24 kHz audio that stays coherent for several minutes, and can turn a whistled tune into a style from a description; the MusicCaps set of 5.5 thousand music and text pairs was published with the model. MusicGen (Copet et al., 2023) works with a single language transformer thanks to token interleaving schemes across several codec streams, and it accepts a text or melody condition. Diffusion models in the latent space of audio are developing in parallel. Quality is measured with the FAD metric and by listening, and a separate task is searching for similarity with training tracks.
Example
A coffee shop chain orders background music for its halls. A marketer generates several hours of calm tracks in different styles on a plan with a commercial license and keeps the service terms together with the files.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map