search

cmuratori / moriarty_audio

Open request for help filtering "lost" audio of an interview session with Professor Brian Moriarty.
8 stars 0 forks source link

Has This Been Resolved? #5

Open mathaou opened 1 year ago

mathaou commented 1 year ago

Hello!

I was just wondering if this skipping issue had been resolved or if additional passes would be beneficial. I have some ideas that I think would really make it a lot less noticeable or remove it entirely.

cmuratori commented 1 year ago

It never really got resolved to "good". I ended up using a contributed version run through Izodope from Thekla's audio people. It's not great, but it's better than the original (certainly).

mathaou commented 1 year ago

It never really got resolved to "good". I ended up using a contributed version run through Izodope from Thekla's audio people. It's not great, but it's better than the original (certainly).

  • Casey

So is this already in the pipeline for release or would one final pie-in-the-sky-ultimate-solution be worthwhile?

No promises, but I think I have some ideas that other people haven't tried yet.

A L G O R I T H M S

cmuratori commented 1 year ago

Really, it can be "solved" at any time if someone has the wherewithal to solve it :) I can always go back and replace the source audio with an improved source audio and re-render things.

I do think we currently have the technology to make a pristine version here, because, for example, something like ElevenLabs clearly suggests we could use clean samples of Brian Moriarty talking (which I have a lot of) to remake the bad versions, etc. But unfortunately I don't think anyone has actually made a "rerender voices" application yet, so, it may be the kind of thing where we have to wait until someone does (I don't have any time to investigate such a thing myself).

- Casey

mathaou commented 1 year ago

I work with real-time networked audio and there are some trade secrets for packet loss concealment that might come in handy here. Also, the weird blips in the Moriarty clips generally sound to be about the same every time, and often times it happens at the same interval so I was thinking of a rather naive approach where I create some sort of "mask", detect when the blips occur, FFT those instances and subtract the waveforms that match the mask. I'll have to run that one by our DSP engineer to see if that makes sense at all though.

cmuratori commented 1 year ago

Well, if you make any progress, definitely let me know!

Thanks, - Casey