How an AI DJ actually works (and what it can't do)
An AI DJ is not one model. It is analysis, a planner and an audio engine. Here is each piece, and where the AI really matters.
When people hear "AI DJ" they picture a model that listens to two songs and mixes them. That is not how it works, and the real design is more interesting.
An AI DJ is three layers: analysis, a planner and an engine. Only one of them needs a language model.
Layer 1: analysis
Before a song can be mixed, the system needs facts about it. These come from signal processing, not AI:
- Tempo and beat grid. Where every beat and every bar's first beat falls.
- Key. Usually expressed on the Camelot wheel, so two songs can be checked for harmonic compatibility.
- Energy over time. How loud and dense the song is at each second.
- Cue points. Where the intro ends and where the outro begins.
- The hook. The section that is both loud and repeated.
Everplay does this once per song, ahead of time, and stores the result with the track. Nothing is analysed while you listen.
Layer 2: the planner
The planner decides what to do with a pair of songs: which exit, which entry, which transition, how long. A good planner is mostly rules: do not blend songs in clashing keys, do not beat-match across a 30 percent tempo gap, prefer a cut over a bad blend.
This is where a language model helps. It can read the facts about both songs and choose among the moves with a reason attached, and it can map a loose request such as "singable road trip" onto a known vibe. Heuristic fallbacks matter just as much: when the model is slow or unavailable, a rule-based planner takes over so the music never stops.
Layer 3: the engine
The engine plays the audio. It runs two decks and schedules every change on the audio clock rather than with timers, because timers drift and a drop that is 40 milliseconds late sounds wrong. A transition is a handful of scheduled gain, filter and echo changes at exact beats.
What an AI DJ cannot do
- It cannot fix a bad pair. If two songs have nothing in common, the best move is a clean cut.
- It cannot hear the room. A human DJ watches the crowd. Software only knows what you tell it.
- It is limited by its catalogue. Analysis has to exist before a song can be mixed.
The one rule that matters
The magic of a mix dies with latency. If the system visibly thinks at the moment of the transition, the illusion breaks. So the design rule is: prepare everything ahead, and let the user's action only execute a decision already made.
Want to hear the result? Open Everplay, or read about how it works.