5 ms·

OCRing Music from YouTube with Common Lisp

63 points by superdisk 2y ago

kanwisher 2y ago
honestly this would be better with an AI model
- teruakohatu 2y ago
  > honestly this would be better with an AI model In the article the author tried Tesseract which uses ML and has some neural network models, and also tried ChatGPT. I have come to the same conclusion as the author when doing OCR that needed 100% accuracy. When you know the font, spacing and the layout is fixed, old school statistical analysis of the pixels works a treat.
  - register 2y ago
    Completely second that. This is my experience as well.
  - Vampiero 2y ago
    You can generalize that to anything: when you know the problem domain so well, why the hell are you using ChatGPT to solve any problem within it? Use the most specialized tool for the job or you're just wasting CPU and memory (and electricity, and money, and time). Same goes for a neural net trained on every possible character set. If you know the font and character size in advance it's way overkill. It's a bit more effort to set up since you actually have to set it up. But at least it's done right.
- secondplacetho 2y ago
  ML is the second best answer to everything, and very rarely the first best answer. Of course it'd be better than something that is intentionally limiting itself. But that says nothing.
- curt15 2y ago
  By "AI model" do you mean neural nets? "AI" or "ML" are just buzzwords that conveys no real meaning about the underlying mathematics. The underlying models could be something as basic as linear or logistic regression, which depending on the application could actually be more appropriate that full-blown neural nets.
rcarmo 2y ago
Holy cow.
varjag 2y ago
If you're also getting a 500: https://web.archive.org/web/20250106075631/https://nickfa.ro/wiki/OCRing_Music_from_YouTube_with_Common_Lisp https://web.archive.org/web/20250106075631/https://nickfa.ro...
- superdisk 2y ago
  I just restarted the webserver. It's running on OpenBSD HTTPd + MediaWiki + SQLite, and keeping it up has been a perpetual thorn in my side. Oh well. I need to figure out some alternative setup probably.
  - j45 2y ago
    Modify your DNS to put cloudflare or bunny in front of it and you'll be good. Don't stop self-hosting :)
    - zoezoezoezoe 2y ago
      self-hosting means freedom, never stop self-hosting
  - MonkeyClub 2y ago
    Is your VPS on OpenBSD.Amsterdam by any chance? (The 46.23.. address seems familiar.)
    - superdisk 2y ago
      Yep, that's it. The host is (for the most part) fine, but there's either some problem with httpd or the PHP worker pool where it just dies after some number of requests.
      - MonkeyClub 2y ago
        Hi, neighbor! (I'm on server 7.) The service is indeed great, Mischa does an excellent job. Yeah PHP on httpd can be flaky, I'd wish for a lighter solution for wikis.
notpublic 2y ago
Instead of doing a diff, curious if Normalized compression distance (NCD)[1] will yield a better result. It is very simple algorithm: to compare two images, i1 and i2 l1 = length(gzip(i1)) l2 = length(gzip(i2)) l12 = length(gzip(concatenate(i1, i2)) ncd = (l12 - min(l1, l2))/max(l1, l2) Here is a nice article where I found out about this long ago. https://yieldthought.com/post/95722882055/machine-learning-teaches-me-how-to-write-better-ai https://yieldthought.com/post/95722882055/machine-learning-t... From the article: "Basically it states that the degree of similarity between two objects can be approximated by the degree to which you can better compress them by concatenating them into one object rather than compressing them individually." [1] https://en.wikipedia.org/wiki/Normalized_compression_distance https://en.wikipedia.org/wiki/Normalized_compression_distanc...
- johnisgood 2y ago
  Oh interesting, I remember comparing images before, I think I was doing a diff as well, so I suppose this would have worked? Nice to know! They were very small images though. It probably would have added the overhead from compression which in my case would have been detrimental.
  - namvdo 2y ago
    [dead]
  - notpublic 2y ago
    Do try it. We use it for text search in one of our apps and works remarkably well. Basically to find which chunks contain the given text. Since the text can span multiple chunks, a simple string search will not work.
xenonite 2y ago
To OCR music scores, see e.g., https://digitalcollection.zhaw.ch/items/276365b9-0a20-4286-af62-060d70a04402 https://digitalcollection.zhaw.ch/items/276365b9-0a20-4286-a...

63 points by superdisk 2y ago

kanwisher 2y ago

honestly this would be better with an AI model

teruakohatu 2y ago
> honestly this would be better with an AI model In the article the author tried Tesseract which uses ML and has some neural network models, and also tried ChatGPT. I have come to the same conclusion as the author when doing OCR that needed 100% accuracy. When you know the font, spacing and the layout is fixed, old school statistical analysis of the pixels works a treat.
- register 2y ago
  Completely second that. This is my experience as well.
- Vampiero 2y ago
  You can generalize that to anything: when you know the problem domain so well, why the hell are you using ChatGPT to solve any problem within it? Use the most specialized tool for the job or you're just wasting CPU and memory (and electricity, and money, and time). Same goes for a neural net trained on every possible character set. If you know the font and character size in advance it's way overkill. It's a bit more effort to set up since you actually have to set it up. But at least it's done right.
secondplacetho 2y ago
ML is the second best answer to everything, and very rarely the first best answer. Of course it'd be better than something that is intentionally limiting itself. But that says nothing.
curt15 2y ago
By "AI model" do you mean neural nets? "AI" or "ML" are just buzzwords that conveys no real meaning about the underlying mathematics. The underlying models could be something as basic as linear or logistic regression, which depending on the application could actually be more appropriate that full-blown neural nets.

rcarmo 2y ago

Holy cow.

varjag 2y ago

If you're also getting a 500: https://web.archive.org/web/20250106075631/https://nickfa.ro/wiki/OCRing_Music_from_YouTube_with_Common_Lisp https://web.archive.org/web/20250106075631/https://nickfa.ro...

superdisk 2y ago
I just restarted the webserver. It's running on OpenBSD HTTPd + MediaWiki + SQLite, and keeping it up has been a perpetual thorn in my side. Oh well. I need to figure out some alternative setup probably.
- j45 2y ago
  Modify your DNS to put cloudflare or bunny in front of it and you'll be good. Don't stop self-hosting :)
  - zoezoezoezoe 2y ago
    self-hosting means freedom, never stop self-hosting
- MonkeyClub 2y ago
  Is your VPS on OpenBSD.Amsterdam by any chance? (The 46.23.. address seems familiar.)
  - superdisk 2y ago
    Yep, that's it. The host is (for the most part) fine, but there's either some problem with httpd or the PHP worker pool where it just dies after some number of requests.
    - MonkeyClub 2y ago
      Hi, neighbor! (I'm on server 7.) The service is indeed great, Mischa does an excellent job. Yeah PHP on httpd can be flaky, I'd wish for a lighter solution for wikis.

notpublic 2y ago

Instead of doing a diff, curious if Normalized compression distance (NCD)[1] will yield a better result. It is very simple algorithm: to compare two images, i1 and i2 l1 = length(gzip(i1)) l2 = length(gzip(i2)) l12 = length(gzip(concatenate(i1, i2)) ncd = (l12 - min(l1, l2))/max(l1, l2) Here is a nice article where I found out about this long ago. https://yieldthought.com/post/95722882055/machine-learning-teaches-me-how-to-write-better-ai https://yieldthought.com/post/95722882055/machine-learning-t... From the article: "Basically it states that the degree of similarity between two objects can be approximated by the degree to which you can better compress them by concatenating them into one object rather than compressing them individually." [1] https://en.wikipedia.org/wiki/Normalized_compression_distance https://en.wikipedia.org/wiki/Normalized_compression_distanc...

johnisgood 2y ago
Oh interesting, I remember comparing images before, I think I was doing a diff as well, so I suppose this would have worked? Nice to know! They were very small images though. It probably would have added the overhead from compression which in my case would have been detrimental.
- namvdo 2y ago
  [dead]
- notpublic 2y ago
  Do try it. We use it for text search in one of our apps and works remarkably well. Basically to find which chunks contain the given text. Since the text can span multiple chunks, a simple string search will not work.

xenonite 2y ago

To OCR music scores, see e.g., https://digitalcollection.zhaw.ch/items/276365b9-0a20-4286-af62-060d70a04402 https://digitalcollection.zhaw.ch/items/276365b9-0a20-4286-a...