Make terrible-quality MP3s sound better than CD quality with AI remastering
Sep 17, 2017 at 12:57 PM Thread Starter Post #1 of 6

DiscoProJoe

100+ Head-Fier
Joined
Mar 8, 2015
Posts
433
Likes
108
Location
Chongqing, China
Just had this thought go through my head today: it'd sure be nice if you could buy a portable music player (or have software on your computer, or a smartphone app) with an AI program that could take crappy-sounding MP3s, analyze the music data, make "smart" guesses about what sonic data is missing or distorted, and use those guesses to fill in the missing details in the music while you're playing the song!

The AI also could take monaural music files and "figure out" how to separate the sound into two stereo channels (or even 7.1 channels) in a way that sounds good, and play it back that way.

If you want, the program could also extract these modified "smartened" music files to your PC desktop, where you could save them as MP3s, FLACs, WAVs, or whatever, at your bitrate and sample rate of choice.

If these AIs become powerful enough further into the future, then eventually you could take a 384 kHz / 32-bit, 7.1-channel FLAC file, convert it all the way down to a telephone-quality MP3 at 16 kbps, 11 kHz, mono, and then play this bad MP3 using the AI program on your music player to make it sound >99% the same as the original FLAC file -- through all 7.1 channels!

Wouldn't that be something?

It gets even better. Play song recordings from the early 20th Century. Flapper music, swing music, whatever. If you have a nice sound system or good headphone system, then these songs will sound like you're witnessing the show live at a concert.

Play pop-song recordings from the early 21st Century with crushed dynamics and "loudness war" shenanigans, and they'll also have a live sound with amazing dynamic detail as well.

But, of course, this AI remastering wouldn't be limited to just music. How about movies, photographs, and even artwork?

You could play a black-and-white, 4:3-aspect-ratio movie from the 1930s, and it would look and feel as if it were filmed today with 4K-resolution video equipment, 32-bit color, 16:9 aspect ratio, and with highly-detailed 7.1-channel surround sound. (Or could even be made for a virtual reality (VR) headset.)

Old photos from the 1800s would appear as if they were taken today with a professional 20-megapixel digital camera. Famous historic artwork could be rendered to look like real photographs. And, of course, cartoons and animated films would look like they were filmed with live characters and real scenery.

I couldn't find much information online about this topic, but did manage to turn up this article about old films, photos, and drawings.

Anyway, I was just wondering if anyone else here on Head-Fi has ever had these thoughts go through their mind. Every time I see or hear about a portable music player that costs a ridiculous $3,000, I've started thinking to myself, "Hmmm....Maybe it has one of those futuristic AI remastering apps that can make a telephone-quality MP3 sound like 7th heaven!"

Oh well,...it's common knowledge that you can take high-quality audio and video, and reduce it down through compression. But most people believe it's impossible to do the opposite.

In the future with AI technology, though, this "opposite" almost certainly will be possible. Can't wait!

How about you?
 
Last edited:
Sep 18, 2017 at 12:12 PM Post #2 of 6

DiscoProJoe

100+ Head-Fier
Joined
Mar 8, 2015
Posts
433
Likes
108
Location
Chongqing, China
Here's one way that such an AI program could be developed for music: feed it hundreds of thousands (or millions) of high-definition FLAC files of song recordings from the last 35 years, spanning every music genre and sub-genre possible. (These FLAC files would need to be genuine ones, and not fake FLACs converted from MP3s.)

When reading each music file, the AI program would train itself. First, it would create a temporary telephone-quality music file from the FLAC, equivalent to an MP3 at 16 kbps, 11 kHz, mono. Then, the AI would compare the music data from this bad-sounding temp file...to the original FLAC file. The AI would think, "OK, if I ever encounter a file that sounds like this, then I need to make it sound like that."

Finally, the AI would delete the bad-sounding temp file, and move on to the next FLAC track and repeat the process.

After its training from a million examples of songs, and from how to take each one from extremely low quality to extremely high quality, then hopefully, this AI program would know what to do when encountering bad-sounding MP3s! The results may not be 100-percent perfect, but as the technology improves and as the AIs get more intelligent, the output will be closer and closer to the original.

Any thoughts on this?
 
Sep 18, 2017 at 1:08 PM Post #3 of 6

eaglesgift

500+ Head-Fier
Joined
May 19, 2017
Posts
549
Likes
426
Location
Chiang Mai, Thailand
I thought the Sony players already did that with mp3s? Not AI but making them sound better. I can't see anybody investing much time in such a project though: storage is getting bigger and cheaper all the time so it's just not necessary.
 
Last edited:
Sep 18, 2017 at 3:58 PM Post #4 of 6

DiscoProJoe

100+ Head-Fier
Joined
Mar 8, 2015
Posts
433
Likes
108
Location
Chongqing, China
I thought the Sony players already did that with mp3s? Not AI but making them sound better. I can't see anybody investing much time in such a project though: storage is getting bigger and cheaper all the time so it's just not necessary.

You're not comparing apples to apples. Besides, storage space for music files isn't the main issue with this, either, especially since the program files for an AI app could be very large, anyway.

I'm sure there are plenty of people who would love to hear Louis Armstrong, Glenn Miller, Elvis, The Beatles, and the terribly-mastered Imagine Dragons album from a few years ago sound incredible. How about all those rare tracks in your music collection that have sub-par recording / mastering quality? Don't you ever wish they could sound amazing?
 
Sep 18, 2017 at 4:49 PM Post #5 of 6

eaglesgift

500+ Head-Fier
Joined
May 19, 2017
Posts
549
Likes
426
Location
Chiang Mai, Thailand
You're not comparing apples to apples. Besides, storage space for music files isn't the main issue with this, either, especially since the program files for an AI app could be very large, anyway.

I'm sure there are plenty of people who would love to hear Louis Armstrong, Glenn Miller, Elvis, The Beatles, and the terribly-mastered Imagine Dragons album from a few years ago sound incredible. How about all those rare tracks in your music collection that have sub-par recording / mastering quality? Don't you ever wish they could sound amazing?
Yes, it would be great if we could fix some of the badly mastered and recorded music made in the last 100 or so years; I was just giving you my thoughts on your first point, which essentially seemed to be about reconstituting MP3 files and I don't think there's a need for that. I'm not sure what AI could do for poor recordings and whether people would question the 'authenticity' of the results but I'd definitely be interested to find out.
 
Sep 18, 2017 at 4:59 PM Post #6 of 6

Kira69

Head-Fier
Joined
Dec 28, 2014
Posts
89
Likes
69
Audio restoration, my favourite topic.

Sadly something like "reconstruct missing information with accuracy" when it's already lost it's far from possible. At least for the next decades.

Today we already have tools that use that kind of algorithms. On audio we have Stereo tool, iZotope RX, etc. For images we have waifu2x for example. For video we have Twixtor to increase framerate. And many more.

But DiscoProJoe is right. A software doing what he said didn't exist yet afaik. And it could be a big step in this realm.
 

Users who are viewing this thread

Top