Flutist vs. Neural Networks

🎼 A flutist put Suno and AssemblyAI through a listening test

On Habr, in the blog of the company Bothub, a hands-on breakdown of music neural networks was published — article 1082846. The author is a flutist, that is, a practicing musician, and the breakdown was done by ear, not by metrics.

The mechanics are simple: two models with different tasks. Suno composes music from a text prompt, AssemblyAI works on the receiving end — it recognizes audio and vocals. The author requested Suno to generate in the key of E minor, and for AssemblyAI she sang herself — testing how the model hears and recognizes vocals and notes.

The format is a point-by-point account: where the models understood the task and where they "played their own thing." The theme is stated directly in the abstract: what AI music "can already do" and what it "still imitates."

If you strip away the keys, the picture is this: Suno and AssemblyAI sit at two ends of the same pipeline — one turns text into sound, the other turns sound back into analysis. And at both ends, the models don't always understand the task accurately: in some cases the generation drifts from the prompt, and AI music from text still partially imitates the intent rather than reproducing it.

The flutist described this experiment in full in the Bothub article on habr.com