Since I work on multimedia stuff, I had some chances to look at different specifications for audio codecs as well as video ones. And comparing those specifications led me to rather surprising conclusion:
Video codec specifications are better than audio ones
I admit that I might miss some details or interpret something wrong yet here’re my impressions.
Video codec specifications tend to be complete and while they are not always easy to comprehend you can write codecs after them (well, maybe in VP8 case you need to add some glue code to reference decoder disguised as specification). And those specs usually cover everything, including extensions and rather freely obtainable (usually drafts are good enough for all practical purposes). I know mostly ITU H.26x, MPEG video and SMPTE VC-1 codecs but that seems to apply to all of them.
Audio codec specifications often lack those features. Looks like they offer you more or less complete version of core decoder and good luck finding description of extensions. And even then they manage to lie or omit some things.
MPEG Audio Layers I-III — no objections (except for Layer 3 bitstream design, of course).
AAC — nothing wrong with core bitstream parsing. Heck, I even managed to write something that produces valid AAC stream (sometimes) that can be decoded into recognisable sound (with a bit of luck). Now look at the extensions — they are very poorly documented. That was the reason why FFmpeg got AAC SBR and AAC PS support that late.
ATSC A/52 (some call it AC3) — again, nothing wrong with core decoder. FFmpeg even got encoder for it long before native decoder (you can blame liba52 for that too). But E-AC-3 is a different beast. It’s mostly okay unless you try implementing enhanced coupling or dependent streams. The former has some bugs so implementing it right by the specification will result in decoder failing to parse bitstream sometimes and the latter is almost fine but in version available in the wild there’s no mention that some of extension channels are actually channel pairs. Good luck figuring it out by yourself.
DCA — it was fun when I discovered that actual CRC calculation is performed with polynomial different from the one given in specification. Luckily nobody bothers about it. Some of the tables are not given in specification — DCA is the codec with the largest amount of tables if you count its extensions (TwinVQ is a close competitor though), so they decided not to give tables with more than 1024 elements in specification. You need reference code for that. And good luck finding specifications for extensions. And I assure you that like in case with E-AC-3 reference decoder sometimes does different things than written in spec. The most wonderful part? Your decoder should be bitexact if you want lossless mode operating properly and looks like the only way to do that is to copy a lot of stuff from reference decoder.
Speech codecs (AMR-NB, AMR-WB, ITU G.72x, RFC xxxx) — some of them nice but I still have that impression that most of them had specifications written by the next formula: “Here’s how it should operate in principle, I don’t remember exact detail anyway but we need to write something here, you have your reference decoder source so bugger off”. I remember looking at
those G.72x specs (some are quite nice though), I remember troubles Robert Swain had with AMR-NB decoder (I tried to help him a bit with it too) and there’s some speech codec (iLBC?) that simply dumped all source code files into RFC body without much explanation.
That’s why I claim that audio codec specifications are generally worse than video codec specs. I think if somebody ran simple test on specs assigning penalty points for things like:
- containing non-obvious pseudocode constructs like
QTMODE.ppq[nQSelect]->inverseQ(InputFrame, TMODE[ch][n]) - containing five-line pseudocode for what can be expressed in one clear sentence
- containing source code listings (bonus points for spanning more than one page)
- omitting details essential for implementation
- lying about details (i.e. when reference decoder and specification disagree)
- assigning decoder to do irrelevant tasks (like upsampling, postfiltering and such)
virtually no audio codec would have zero score unlike video codecs.
Even when working on the AAC channel coupling element which is part of AAC-LC and has been around since MPEG-2, I found errors in the spec that were fixed in 13818-7:2006/Cor.1:2009. That is more than a decade after the original publication of the specification.
There are still a ton of tiny errors that never got fixed.
Maybe I’ll do a blog post about all known outstanding AAC specbugs.
@Alex: Or start a new MultimediaWiki page. That might inspire others to add. If Google doesn’t update the VP8 spec soon with all the corrections they’ve received, I’m planning to do the same for VP8.
You have a blog? Awesome! The more multimedia blogs, the merrier!
I haven’t read any video codec spec, but I’m willing to take your word that they are better written. My guess (and I’m just guessing) is that the reason for that would be about complexity (of code, not CPU). Because video is less mature than audio and it has to deal with huge amounts of data, the number of operations you can do on each pixel is pretty limited (dozens maybe). For audio, we can easily afford *thousands* of operations per sample, if not tens of thousands. This means there would be more operations to describe for audio than for video. Another factor could be related to prediction. In video the prediction (for P frames) is technically borderline-unstable (i.e. non-leaky integrator), which means unless your decoder spec is very precise and bit-exact, you’ll have drift and crappy output. On the audio side, codecs either don’t have prediction or when they do, the prediction is leaky and this clearly stable.