A generative music model can produce a convincing minute of audio from a sentence. That does not make it a musician that can hear the note you just played and answer without flattening the soloist.
MazzikaAI works in that gap. The system presented by Jiaxin Du, Boulbaba Abdeljaouad, Yong Zhuang and Haoyu Li does not retrain the generator. It builds a rule-based layer that continuously turns a musician's performance into new prompts for Google Lyria RealTime.1
The foundation model stays general. The musical knowledge lives in front of it.
A compiler between the keyboard and the model
A musician plays a MIDI keyboard in the browser. The system captures notes, sustain, gestures and voice commands. A backend maintains a working state describing register, density, movement, dynamics, mode and other useful features.1
That state is interpreted through an explicit knowledge base for Arabic maqam. The authors document six core maqamat, microtonal intervals, characteristic ornaments and ensemble rules.1
The controller then compiles the current state into text. The prompt tells the generator what to play now, which instruments to use, what role to take and how to answer the soloist's phrase. Lyria streams the audio, and the loop continues.1
“Compiler” is the useful word here. The musician never writes the prompt. Their playing is the source language.
Keeping expertise outside the model
The common response to a model that poorly represents a practice is to find more data or fine-tune it. MazzikaAI takes another route.
The paper starts from the mismatch between mainstream generative music systems and Arabic maqam, where microtonal intervals and modal behaviour do not fit neatly into Western equal-tempered assumptions.1
Instead of hoping Lyria already knows how to behave, the controller encodes that knowledge in an inspectable form. One rule can ground a microtonal degree. Another can constrain instrumentation. Another can request an answering phrase that echoes part of what the soloist just played.1
Architecturally, this is almost old-fashioned: knowledge base, working memory, deterministic rules. The unusual part is the final actuator. Instead of driving a symbolic synthesizer, those rules drive a large audio model through natural language.1
Prompting as a control protocol
The idea travels beyond music.
We usually describe a prompt as something a person writes. Here, nobody wants to type “play softer, stay in bayati, answer this phrase, remove the drums” every two seconds. Text becomes an intermediate protocol generated by one machine to steer another.
That separates two jobs. The foundation model supplies sonic capability. The controller decides what is acceptable in the musical context.
The authors report sub-second key-to-audible-update latency, expert playing sessions and ablations around maqam grounding, instrument suppression and re-prompt gating.1
The evidence has limits. This is a fresh preprint evaluated by its authors. The claim that the architecture can generalise to other musical traditions is a proposed direction, not an established result.1
Still, the mechanism is useful. When a general model lacks domain culture, changing its weights is not the only option. You can put a small explicit system in front of it that knows what it wants.
In MazzikaAI, the generative model is not the knowledgeable musician. It is a very capable instrument being kept on course by a much smaller set of rules.