Hacker News Reader

I think the first step would be to be able to convert a lossy format as MP3 or ogg to a lossless format as FLAC or Wav. Being able to retrieve close in off lost data from lossy compression would be impressive.

View on HN

MP3s generally have a shelf around 16-18khz. The rest of the audible spectrum data is impossible to retrieve if it doesn't exist. This is why transcoding from a lossy->lossless format is a bad idea.

Yes, I know and that's what would be more impressive for me to see a ML/Algo that could recreate close enough data distribution from lossy to lossless. Not exact reconstruction but close enough in the possible area (so < 16-18kHz). It may or may not be possible but it is more akin to take the inverse of the model of degradation of MP3 (it could be totally impossible).

Sounds like you're describing DLSS [0], but for audio.

[0] https://en.wikipedia.org/wiki/Deep_learning_super_sampling

Reply on news.ycombinator.com