Generate a rap with AI and see if it can fool your ears
I compared Fenix Flexin's "Rubberz" with a mimicry track generated by Suno, actually running through the complete zero-to-one process. Honestly, after finishing it, I felt a bit dazed—not marveling at how strong AI is, but realizing that the question of "does it sound human" has technically become blurred.
Background first: Fenix Flexin, a member of the LA rap group Shoreline Mafia, recently saw his single "Rubberz" climb to #58 on the Billboard Hot 100, only to face accusations of being AI-generated. I listened to the song. The flow is smooth, enunciation is a bit electric, and the beat is typical West Coast trap. Does it sound like AI? Yes. But does it not sound like a real person? Not entirely. This gray area is what's worth discussing today.
The tutorial has three parts: generating a rap, adjusting it to sound more "human," and identifying where it might give itself away. All done using Suno (currently the mainstream AI music generator, free web tier is sufficient). If you use Udio, steps are similar.
1. Register and Enter Creation Interface
Go to Suno's official site (suno.ai), click Sign up, and log in with Google or Discord. Free users get 50 credits daily; generating one song costs 5 credits, enough for 10 tries.
After logging in, click Create in the top left corner of the homepage to enter the creation panel. You'll see two input boxes: Song Description and Lyrics. Here's the key part.
2. Write Prompt, Generate First Version
First, write a description, format reference:
Style: West Coast trap, 808 bass, heavy hi-hats, male vocal with autotune, aggressive flow
Theme: Bragging about wealth and street life, with a sarcastic tone
Structure: intro + verse 1 + chorus + verse 2 + chorus + outro
Note: Do not write words like "AI generated" or "fake"; the system tends to generate more mechanical stuff. I fell into this trap the first time; the result sounded like electronic synthesis reading lyrics aloud, with no groove at all.
I casually wrote some lyrics mimicking typical West Coast content:
Paste these lyrics into the Lyrics box, and put the description above in the Description box. Click Generate and wait about 30 seconds.
3. Listen to First Version, Find Issues
After the first generation, you'll hear a snippet from 30 seconds to 1 minute. My result: Beat was okay, 808 drums present, but vocals were very fake—every syllable too clear, no breathing sense, no improvised "dragging" or "rushing" beats, flow sounded like a robot chanting.
This is AI's most obvious flaw: Rhythm is too regular. Real rappers intentionally drag certain words or suddenly speed up at sentence ends; AI doesn't do this.
4. Adjust Prompt, Fix the "Fake" Issue
Back to the Description box, add specific instructions to break the regularity:
You can also specify reference tracks (if Suno upgraded relevant features, but current version doesn't support direct referencing, so use text descriptions). Additionally, manually add placeholders in Lyrics, like (breath) or (ad-lib: yeah) at sentence ends.
Regenerate. This time results were much better: Vocals started having slight sway, swallowed sounds at sentence ends, even appeared "uh" improvisations. But it still lacked "emotion"—anger, sarcasm, excitement. These emotions are expressed through changes in enunciation force in real performances; AI currently struggles to simulate this.
5. Pitfall Section: Three Most Overlooked Details
Pitfall 1: Lyrics too neat. My first version had equal character counts per line and perfect rhymes. Result: AI flow sounded like keeping time. Real rappers intentionally mix long and short lines, e.g., 8 chars in line 1, 11 in line 2, inserting a mouth movement like "tsk" in between. Suggestion: Intentionally leave one or two non-rhyming words when writing lyrics, or add a "pause" marker.
Pitfall 2: Writing "autotune" in description hurt it. I initially wrote "with autotune," resulting in fully electronic vocals, sounding like a robot. Later changed to "natural vocal, slightly processed," closer to real studio processing—slight pitch correction, but not fully automated.
Pitfall 3: Generation length. Free version maxes out at 2 minutes, but a full rap is usually 3-4 minutes. You can use Extend to continue, but extended parts often suddenly change key or rhythm messes up. Suggestion: Write a short song under 2 minutes first, don't be greedy.
6. How to Judge if Others Are Using AI
After making my own AI song, I went back to listen to "Rubberz" and found it hard to judge. Fenix Flexin's song has some "human" clues: e.g., a distinct breath intake in the chorus, verse 2 rhythm suddenly slows by a beat, seeming like improvisation. These are hard for AI to mimic currently.
Conversely, AI songs might have "fake humanity"—e.g., deliberately added breathing sounds, but timed too regularly, once every 4 bars, like an alarm clock. Real breathing is random; sometimes singers hold breath during inhale moments to finish a phrase.
7. Next Steps After Learning This
If you're interested in AI music, don't stop at generation. Try:
- Generate a song with AI, then record your own vocals over it to see if you can pass it off
- Compare Suno and Udio generation quality, see which excels in which style (I tried; Suno's trap is good, Udio's R&B is better)
- Research Spotify's AI detection algorithm: They removed 75 million AI songs last year. How did they do it? There are public reverse engineering analyses online.
Physix Frontier