It helps to separate the question into four smaller ones: what does the model actually predict, what does your prompt control, where do the vocals come from, and who does the mixing. None of it is magic, but it also is not simple playback of pre-recorded loops.
The model: predicting audio, not searching for it
A music generation model is trained on a very large amount of audio paired with descriptions, so it learns statistical relationships between words like 'slow acoustic ballad' and the acoustic patterns those words tend to describe — tempo, instrumentation, chord movement, vocal tone. When you submit a prompt, the model does not look up a matching song and return it; it generates a new waveform, section by section, that fits the description you gave it.
Autunes runs this on the Zori model family. Zori5 Turbo is tuned for speed and is available on the free tier; Zori5 XL trades some generation speed for higher fidelity and is available from the Pro plan up. Both work the same way conceptually — they differ in size and training emphasis, not in method.
The prompt: your only steering wheel
Your text prompt (genre, mood, instrumentation, era, vocal style) and your lyrics are the two main inputs the model conditions on. Structure tags in the lyrics — [Verse], [Chorus], [Bridge] — tell the model where to change energy and melody, similar to how a producer would mark up a lead sheet. Vague prompts get vague, generic results; specific prompts get specific results, which is why prompt-writing is its own skill (worth a separate deep-dive).
Vocals: sung, not spoken-then-pitched
The vocal track is generated as singing directly — the model has learned pitch, vibrato, breath and phrasing as part of the same audio-prediction task, not as a separate text-to-speech step glued onto a beat. That is why a good AI vocal has natural-sounding dynamics rather than the flat, robotic quality of older speech synthesis. It is also why unusual or foreign words can occasionally be mispronounced — the model is predicting sound, and it is only as good as the pronunciation patterns it has seen.
The mix: baked in, then editable
Level balancing, EQ and effects are generated as part of the same pass, so a finished AI song already sounds mixed rather than arriving as raw, unbalanced stems. If you want to change that balance afterward, Autunes' Studio splits a finished song into six stems (vocals, drums, bass, guitar, piano, other) so you can remix, mute or rebalance individual parts, export MIDI, or hand stems to a DAW for a proper manual mix.