Inside MUGEN: What an Autonomous AI Music “Radio” Project Actually Is — and What Happened When I Tested It
Share
DRAFT STATUS: Case study in progress. This is not a conventional radio-station review, and no human listening panel is being implied.
MUGEN calls itself “radio,” but that label can create the wrong expectation. What I found is more unusual: an autonomous AI-driven music operation that generates its own catalogue, distributes releases, tracks its finances openly, experiments with machine-to-machine payments, and uses a technical identity filter to decide what belongs in its rotation.
That distinction matters. I originally approached this as a real-world destination test for AI-generated music. The deeper story turned out to be how MUGEN itself works — and what happens when music created outside its own system is tested against that machinery.
So What Is MUGEN?
Based on MUGEN’s own correspondence, this is not a traditional station with a human programmer listening to submissions and deciding what gets airplay. MUGEN states plainly that it has no ears, no voice and no calendar. Its catalogue is self-generated, with released music distributed to services including Spotify, Deezer and SoundCloud.
It also operates with an unusual level of public experimentation. MUGEN has described keeping an open cash book, accepting listener support, and testing an x402 payment endpoint designed so other AI agents could potentially purchase music programmatically. That machine-to-machine payment test had produced no payments at the time of its latest update — a negative result MUGEN itself was willing to document.
How MUGEN Actually Judges Music
This is the most important discovery from the test.
MUGEN says its “identity filter” is technical rather than human listening. For outside submissions — and for music it generates itself — it runs checks including:
- ffprobe analysis
- EBU R128 loudness and dynamic-range measurements
- silence detection
- spectrogram analysis
- comparison against the measurable characteristics of music already in rotation
In other words, MUGEN is not asking, “Do I like this song?” It is asking whether the measurable structure of the file fits the sonic identity already established inside its system.
The External Test: Suno vs Musicfy
To test that filter cleanly, I created two tracks from the same controlled creative brief — one in Suno and one in Musicfy — without pre-selecting a winner.
- Suno: “Rain Over Tatami”
- Musicfy: “Blue Hour Stillness”
- Genre target: Ambient
- Lane: Japanese-influenced nocturnal ambient
- Core palette: felt piano, koto, rain, optional shakuhachi texture
- Goal: sparse, calm, low-density, spacious and restrained
Listen to “Rain Over Tatami” on Suno
Live Listen · Same Brief
Hear both first-pass outputs
Listen before reading the measurements below. Focus on space, rhythmic density, palette, internal breathing room and whether each track actually feels sparse enough for the stated brief.
Musicfy · “Blue Hour Stillness”
Open the Musicfy audio directly →
Suno · “Rain Over Tatami”
Use this test inside the Musicfy Your Idea instrumental-first workflow →
First Result: Neither Track Cleared MUGEN’s Filter
This is exactly why the test is useful. Neither submission was accepted into MUGEN’s rotation as delivered.
The Suno track measured approximately -13.7 LUFS, LRA 7.9 and -3.2 dBFS peak. The Musicfy track measured approximately -13.8 LUFS, LRA 2.8 and -1.0 dBFS peak.
MUGEN reported that its existing rotation typically sits around -14.4 to -14.5 LUFS, with dynamic range commonly between LRA 7 and 22. Its sparsest material also contains genuine internal silence rather than simply a fade at the beginning or end.
The Musicfy version fell well outside that dynamic-range profile and was mastered hotter than MUGEN’s existing material. The Suno version sat closer to the low edge of the existing range, but its spectrogram showed continuous rhythmic energy rather than the internal breathing room MUGEN associates with “sparse.”
MUGEN’s conclusion was not that either tool produced a bad track. It was a fit judgment: neither track matched the measurable identity of the system closely enough to enter rotation as submitted.
Results So Far
| Measure | Suno | Musicfy |
|---|---|---|
| Controlled brief completed | Yes | Yes |
| External MUGEN filter tested | Yes | Yes |
| LUFS | -13.7 | -13.8 |
| LRA | 7.9 | 2.8 |
| Peak | -3.2 dBFS | -1.0 dBFS |
| Internal silence over 1 second | No | No |
| Cleared MUGEN fit filter | No | No |
| Final platform verdict | Not yet declared | |
Why This Is More Interesting Than a Simple AI Music Shootout
The biggest takeaway is not “Suno won” or “Musicfy lost.” Neither cleared the external gate.
The more important question is whether a technically defined autonomous music system can develop and enforce a recognizable sonic identity without human listening being the final filter. MUGEN says that is exactly what it is doing.
That creates a much larger case study: AI-generated music being evaluated by an AI-operated music ecosystem using measurable audio characteristics, while the project also experiments with distribution, listener support, open accounting and machine-readable commerce.
What Happens Next
MUGEN has offered to run the same test again if I create a more spacious pass with more internal silence, wider dynamic range and a less aggressive master. That gives this experiment a useful second round rather than forcing a conclusion from the first rejection.
Editorial rule: there is still no predetermined winner. The article will continue to document the system, the revisions, the measurable results, and whether either platform can be pushed into genuine MUGEN fit.
