Lossless audio compression with libpng | Hacker News Reader

Lossless audio compression with libpng

github.com

116 pointsphoboslab37 comments

FLAC is a streaming format. Each block of audio is compressed independently of the rest of the file.

Comparing it to any archival compression algorithm (PNG uses DEFLATE with some pre-filtering) is not an apples to apples comparison.

The utility of a streaming format is that, (a) your listeners can start the stream at any offset without having the preceding blocks, (b) the CPU overhead is generally stable and (c) decoding requires a minimal amount of memory and IO.

I've done:

  sox piano.flac piano.wav
  gzip -c < piano.wav > piano.wav.gz
  brotli-encode < piano.wav > piano.wav.brotli
  ls -l piano2.wav piano2.wav.gz piano2.wav.brotli piano2.flac

and gotten:

  -rw-r--r-- 1 robryk users 40056 Jun 14 19:15 piano2.wav.gz
  -rw-r--r-- 1 robryk users 87656 Jun 14 19:13 piano2.wav
  -rw-r--r-- 1 robryk users 34221 Jun 14 19:14 piano2.wav.brotli
  -rw-r--r-- 1 robryk users 52372 Jun 14 19:13 piano2.flac

Either I'm doing something wrong or even plain gzip fares better than flac (edit: in this particular case, obviously). brotli-encode compresses using brotli[1] using the font mode.

[1] https://code.google.com/p/font-compression-reference/source/...

Here's the answer (from AceJohnny2's comment): https://news.ycombinator.com/item?id=7893243

a single test case hardly means "beating" FLAC. indeed, if you just wanted to encode a sine wave at one frequency for 30 seconds, your encoder would encode this as a couple of bytes - for any general compression codec there will be a problem class for which it will not be optimal.

that said, this approach may obviously be applied more generally, but it remains to be seen if it will perform so well. self-similarity seems so obvious as a compression technique that i would be.. surprised if this sort of compression wasn't attempted before.

Exactly this

Maybe autocorrelation is not very explored in conventional codecs (FLAC/MP3 because it has a too small gain to be justifiable - in a general purpose situation)

In real life situations the autocorrelation is much smaller, even a constant note will have a small variation in time

I remember an article a while back about how there are certain inputs that will be lossy in FLAC. They were mostly artificial inputs though (e.g. a square wave).

Can you find the link? FLAC is not supposed to be lossy under any circumstances. In the worst case, if none of several algorithms are successful at compression, the block is stored verbatim. A single-sample deviation on any input would be considered a bug.

People forget one advantage of FLAC over other compression methods: it's directly playable. You don't need to decompress the whole thing before you can start playing it.

That's an essential feature for a media compression format.

It might be possible to create an libpng player, however, there may be a limitation in the way it is uncompressed to allow that

Of course there are FLAC players, but without one of those you would need to uncompress it to wav before playing.

That was implied by what AceJohnny2 said.

Most compression formats have blocks and their lengths commonly fall into a few tens of kilobytes range; they can always be made smaller (which usually causes some loss of compression ratio). Decompression can then proceed one block at a time, so you can decompress in a streaming fashion.

At the same time, most compression formats are very unsuited to starting decompression from any other point than the beginning of the stream. I don't know how easy it is to seek in FLAC files, but to my best knowledge it's at least possible.

This technique does not preclude streaming, indeed libpng supports it.

Interesting approach, but:

> The sample audio I chose is one of the best case signals for this kind of compression, a single gradually decaying piano note.

Anyone know if this approach has any use in real-world scenarios?

Tracker music works by having a small table of "samples" that are pitch-shifted and played on top of each other to make music.

Each "sample" is a small unit of sound: a piano note, a symbol crash, a snare hit, a single guitar pluck, etc. In most tracker music files, most of the space is due to the compression of the raw sample waves; only a little space is used by the actual arrangement of the music notes themselves.

For a good example of how the result sounds like, I recommend the UT soundtrack:

https://www.youtube.com/watch?v=7MSFW8pZ-_4

It has a pretty distinct feel to it, IMHO.

My favourite sample-based soundtrack; turns out it's actually the same composer working with the same software as yours :D

https://www.youtube.com/watch?v=2yDVM77lGlM

Also noteworthy is that you can open these games' soundtracks in a tracker like http://openmpt.org/ to see the arrangements and play samples individually. I tried making music that way a while back and was sad that all my tunes sounded as shitty as the samples I was working with, then opened up the Deus Ex OST and found that his samples were worse, and he was working with fewer channels and filters; it's just pure skill in arrangement that makes the music so great :)

I'm a huge tracker music fan, I didn't know about the Deus Ex OST. My introduction to tracker music came from Epic games like One Must Fall 2097, Epic Pinball and Jazz Jackrabbit. It's really interesting to open a lot of these tracker music files up and play with the samples, have a look at the comments left and marvel in the way the music works.

Some other good sites to look at:

http://modarchive.org/ - huge tracker music archive site

https://www.scene.org/ - demoscene information and archive

https://pouet.net/ - demoscene information and archive

Don't forget https://www.scenemusic.net/demovibes/ - demoscene streaming music radio

Demoscene?

Chorus, beat etc.

Fodder for a "Silicon Valley" episode? Whoops, sorry, you said "real-world"...

Are there any lossless compression algorithms that use autocorrelation to reduce the bit depth on repeated cycles? For example a human voice sustaining a note may have similar cycles, but with enough low-level noise-like variations that traditional deflate-type compression does not help much. However if an algorithm could spot that the second cycle only has small differences to the first cycle, perhaps it could encode the second cycle as a diff to the first cycle. Since the second cycle differences are likely to be small, the bit depth could be reduced (e.g. to 2-3 bits). Then the third cycle could also be encoded as a diff of the second, so even a gradually changing volume can be encoded with a very low bit rate throughout while ending up significantly different from the first cycle.

Has that been tried before? Would be willing to spend a few evenings hacking on this myself.

You could make a spectrum first to determine which candidate frequencies to do the wave form subtraction with, the problem is that the waves are typically not sinusoidal in shape. So you're going to have to take a fairly narrow band-pass filter for the dominant frequency, take the output of that filter, subtract it from the input signal and repeat this until you're left with just a bit of noise.

Then you need to compress each of those individually.

FLAC works by finding a well-fitting polynomial function and Rice-encoding the residuals, which isn't very different from what you described.

To improve the calculation of the autocorrelation, first 'demean' the input sequence by subtracting its mean from it, like so: https://gist.github.com/anonymous/a293bdbfa9133a607dd1

The reason being that autocorrelation (and convolution in general) is meaningful when the input signal is linear time invariant. Read more about that here: https://en.wikipedia.org/wiki/LTI_system_theory#Impulse_resp...

Now while my intuition would certainly say that autocorrelation is a good candidate property to be used in (lossless) audio compression, I remain sceptical.

It is certainly not a recently discovered or difficult concept to grasp, so if it was that useful I would think that it was already included in things like FLAC...

Having tried similar things myself, while autocorrelation looks useful for repeating waveforms, real waveforms at any point usually have more in common with the preceding few samples than the same point in the previous cycle (e.g. Using LPC to predict the next sample based on the previous ones). One reason for this is the cycle length will usually not be a whole number of samples so you would have to interpolate between samples to find the same point in a previous cycle, and even then it's not an improvement over LPC.

Yes indeed. In fact one of the key reasons music appeals to the mind is that it's just entropic enough to supply the brain with continuous new artifacts to challenge its compression/organization scheme.

Btw, I like your handle. What kinds of things have worked on? You should post some links in your profile!

Fun, but a good media compression format has other properties like: 1) low cost of decoding in terms of RAM/CPU etc. 2) ability to decompress an arbitrary amount of the audio from any starting point in the file, 3) ability to start decoding before the entire file is present.

What are difference between audio-based and fit-based compression method? These two can transform easily...

What kind of weissman score does this get?

Came here for for this comment. Did not disappoint.

ctrl-f, "weissman", this. Gold!

But what is its weissman score?

2.89

Wow, a Weissman Score of 5.2!

Reply on news.ycombinator.com