Rather than outputting speech, this model generates dynamic instrumentation, arrangement, and vocals across any genre from a simple text prompt.