I'm a bit curious, so I'm wondering how sound is actually generated in these kinds of systems. For example, how do they produce something close to human voice quality? Do they work with sound samples from the training data or is it some kind of synthesis? I'd appreciate it if you could explain in detail!
How does Suno AI generate voices?
👁️ 62 views💬 1 replies❤️ 0 likes
1 Replies
Suno AI's voice generation works, in my opinion, like a pleasant melody created by sampling a DJ's singing voice and mixing those samples with different instruments. For example, when we mix salsa records and add bongo hits, those sounds can be combined with human-like MIDI vocals to produce something that sounds like an original vocal. In other words, it doesn't just copy sound pieces from the dataset—it analyzes vocal cords, breath sounds, and accents to synthesize them.
Think of it this way: if I give Suno AI a sample of my bachata guitar playing, it can take those tones and say, "Let me add a vocal in this style, and throw in some choral-like harmonies," creating a completely new three-part guitar+vocals+choir piece. It captures the tone, rhythmic emphasis, and transitions between instruments from the training data. Similarly, it distinguishes the bombastic vocals of Latin pop and produces accordingly, which I think is why people perceive these sounds as so "real."