I fed ElevenLabs Music a single prompt about our open-source MCP agent framework and got back a complete song: vocals, instrumentation, arrangement, the works. Zero post-processing.
Here's what caught me off guard: the vocal phrasing. Not just the melody, but the micro-timing, breath placement, and emotional inflection. The model placed emphasis on "composable" in a way that actually reinforced the technical meaning. It added vocal runs that felt intentional, not algorithmic.
Technical details that worked:
Prompt structure: [Genre] [Mood] [Key technical terms] [Narrative structure]
Generated: 2:04 track with verse/chorus/bridge structure
Quality: Comparable to demo-level indie recordings
What this means: Voice synthesis was the laggard in generative AI. That's changing rapidly. We're moving from "impressive for AI" to "actually usable in production workflows."
Non-English limitations: I tested it with different languages and hit a wall — very patchy results, nowhere near the English quality. Anyone have experience with non-English lyrics? Curious about phoneme handling across languages.
The gap between human and AI musical performance is shrinking faster than I expected. Worth paying attention to.
