I didn't think this would work at all, and now I'm scared

Jul 1, 2022 · 10:20 PM UTC

26
122
23
820
0
Replying to @Jam2go
Gonna corner the stock market with mine.
1
2
That's a very optimistic projection from the AI
Replying to @Jam2go
btw @Jam2go , have you considered that the quality of reconstructing a sound from an image of Coagula is good enough? it might be part of the low quality output problem too
1
Yeah converting to a spectrogram is incredibly lossy, trying to convert back to sound only gives you a blurry idea of what the sound is. You can still pick apart some general shapes/sweeps to the sound though
1
Replying to @Jam2go
labs.openai.com/s/DsnlOdz3KS… Here's the result in dall e 2

Jordan × DALL·E | Detailed Spectogram image of singing

Created with DALL·E, an AI system by OpenAI

labs.openai.com
1
3
Hmm these look more like they’re trying to create waveforms to me. also lmao at the last one having a microphone in front of the spectrogram like the spectrogram itself is singing
1
more replies
Replying to @Jam2go
well it doesn't really work, does it?
1
20
At this low resolution + converting back to sound it’s mostly going to be noise. But the prompts did change the basic shapes of it. “talking” generated vowel like shapes, “singing” generated more sweeps, and “laughing” generated more choppy sounds
2
20
more replies
Replying to @Jam2go @JonTetrino
But... It didn't work? I lt just produced randomness?
1
3
At this low resolution + converting back to sound it’s mostly going to be noise. But the prompts did change the basic shapes of it. “talking” generated vowel like shapes, “singing” generated more sweeps, and “laughing” generated more choppy sounds
1
3
more replies
Load more