Music & AI Glossary

How fast a track moves, counted in beats each minute. A ballad sits near 70, house lands around 124, drum and bass near 174. Tempo is the single most useful thing to specify in a prompt because it changes the feel more than the instrumentation does.

The home note a piece of music resolves to, plus whether it uses a major or minor scale. Major reads as bright or settled, minor as darker or unresolved. Two tracks in the same key can be layered or mixed together without clashing.

A DJ notation that renames each key as a number plus A or B — 8A, 8B, 9A. Keys next to each other on the wheel sound good together, which turns harmonic mixing into arithmetic instead of theory.

The order and length of a song’s sections — intro, verse, chorus, bridge, outro — and which instruments play in each. Arrangement is what makes a track feel like a song rather than a loop that keeps going.

The part people remember and can sing back. Usually the chorus melody or a short repeated phrase. A track can be well produced and still fail because it has no hook.

The vocal melody and lyrics written over an existing instrumental. In pop production the topline and the track are often written by different people, which is why "send me a topline" is a normal request.

Snapping notes to an exact grid position. Fixes sloppy timing, and overdone it removes the human push and pull that makes a groove feel alive.

Where a sound sits between the left and right speakers. Lead vocals, kick and bass are almost always centred; guitars, keys and backing vocals get pushed outward. This convention is exactly what centre-channel vocal removal exploits.

Generating audio from a written description. You write "slow lo-fi beat with dusty piano and vinyl noise" and get a track back. The description is doing the job a producer brief would.

The written instruction given to a music model. Useful prompts name genre, mood, tempo, instrumentation and whether vocals are wanted. Vague prompts return generic results, because there is nothing to steer toward.

A generation method that starts from noise and repeatedly removes it until structured output remains. Most current audio models work on a spectrogram or a learned latent representation rather than on raw samples, because raw audio has far too many values per second.

A compressed internal representation where similar-sounding music sits close together. Generating in latent space instead of raw audio is what makes a full track possible in seconds rather than hours.

A picture of sound: time along one axis, frequency along the other, brightness for loudness. Several music models generate a spectrogram and convert it to audio afterwards, which lets image techniques apply to sound.

Turning continuous audio into a sequence of discrete symbols a model can predict, the way a language model predicts words. The tokeniser’s quality caps the output’s quality — detail lost there cannot be recovered later.

Anything given to the model to steer generation — a text prompt, a reference clip, a tempo, a set of lyrics. More conditioning means less surprise and more control.

Actually running a trained model to produce output, as opposed to training it. When a generator says "generating…", it is doing inference, and the wait is compute time.

Splitting a finished mix back into parts. Neural separators such as Demucs learn what each instrument sounds like; centre-channel cancellation instead exploits stereo geometry, which is far faster and much less precise.

Reproducing a specific person’s voice from recordings of them. Distinct from generating a voice in a style. Cloning a named artist raises likeness and publicity-rights issues in a growing number of jurisdictions, which is why MusicGenerate does not offer it.

How many times per second audio is measured, in Hz. 44,100 Hz is CD standard and 48,000 Hz is standard for video. Higher rates capture higher frequencies; they do not make a track sound better on their own.

How precisely each sample is stored. 16-bit is release quality, 24-bit gives more headroom while working. Low bit depth shows up as hiss in quiet passages.

WAV and FLAC keep every sample; MP3 and AAC discard information judged inaudible to make files smaller. Lossy encoding is one-way — re-encoding an MP3 compounds the damage rather than restoring it.

A loudness measurement that tracks how loud something actually sounds rather than its peak level. Streaming services normalise to roughly −14 LUFS, so mastering louder than that just gets turned back down.

The gap between the loudest point of a mix and the ceiling where clipping starts. Leaving a few dB of headroom gives mastering somewhere to work.

What happens when a signal exceeds maximum level and its peaks are flattened. Heard as harsh distortion, and once printed into a file it cannot be undone.

Splitting stereo into what the channels share (mid) and what makes them differ (side). Isolating the mid gives roughly the centred content; isolating the side removes it. This is the mechanism behind karaoke-style vocal removal.

When two copies of a signal are inverted relative to each other and sum to silence. A problem when it happens accidentally between microphones, and a tool when applied deliberately to remove centred content.

Metadata stored inside an MP3 — title, artist, album, artwork. Players read these rather than the filename, which is why a correctly named file can still show up as "Unknown Artist".

The visual shape of amplitude over time. Useful for finding silence, transients and section boundaries at a glance; it says nothing about pitch or key.

You pay once, or nothing, and owe no further per-use payment. It does not mean free of copyright, and it does not mean unrestricted — the licence still sets the terms.

Genuinely unowned, which for music almost always means public domain through age. Very little modern music is copyright-free, and the phrase is usually used when royalty-free is what is meant.

Permission to release your own recording of someone else’s song. What every cover version needs, whether performed by a band or generated. It covers the composition, not the original recording.

Permission to synchronise music with visuals — a video, an ad, a game. Separate from the right to distribute the audio by itself, and the licence most creators actually need.

Every song carries two rights: the composition (melody and lyrics, owned by the writer) and the sound recording (the specific performance, owned by whoever made it). A cover creates a new recording but still uses the composition.

YouTube’s automated matching system. It can flag a track you legitimately own if that track was registered by someone else — including cases where identical generated audio was claimed first.

A performing rights organisation — ASCAP, BMI, PRS — that collects performance royalties for songwriters. Relevant once music is played publicly or broadcast rather than only streamed.

An arrangement where whoever commissions the work owns it outright, rather than the person who made it. Common in advertising and games, and worth settling in writing before anything is delivered.