Extract HDCD, How I Can extract HDCD by software? Options

[SCIF]

How I Can extract HDCD by software(EAC, WMP9...)?

[SebastianG]

I reviewed the papers about HDCD you can find on the net again recently. Here's my current view on it:

The dithering/noise shaping part they were advocating is nothing special. In fact, it is inferior to what other programs do (like Foobar or my requant.jar tool) and not provably safe in the sense that it avoids nonlinear distortions. You can't even call it "noise shaping" because it only introduces a bit of colored dither which has more power above 16 kHz -- it doesn't shape quantization noise.

If you decode the peak compression thingy (assuming it has been used) you can restore the upper 5 dB of the dynamic range which is usually rarely used (only for the occasional peaks, that's the idea).

Then there's the reconstruction filter selection. The whole idea of this is at least questionable. (I seriously doubt that this constitutes any subjective improvement!)

HDCD also supports some kind of dynamic range compression for the lower levels. It works like this: Quiet signals can be amplified in steps of 0.5 dB (0.0-7.5 dB = 4 bit side information) smoothly before quantization.

The side informations (whether peak compression is used, which reconstruction filter to use and the compressor gain) is transmitted via the least significant bits with the help of a linear feedback shift register. The only information you can't find on the net is how they actually code this side information. So, you can't implement a proper decoder.

Since the "benefits" of HDCD seem rather small it's -- in my humble opinion -- not worth to decode the stream to something like 24/44 or 24/88. One should rather keep it like it is and/or use a HDCD capable post processing algorithm within your player of choice (which only is supported by Microsoft's Mediaplayer IIRC).

In terms of signal-to-noise ratio (I'm not referring to peak signal to noise ratio), proper HDCD decoding won't help you a thing. The only thing you can improve is restoring the slightly compressed dynamics.

Cheers!
SG

edit: fixed typos

This post has been edited by SebastianG: Jun 28 2007, 11:26

[Christopher Key]

I've had a go at understanding hdcd; see,

http://forum.doom9.org/showthread.php?t=129136

If you decode the peak compression thingy (assuming it has been used) you can restore the upper 5 dB of the dynamic range which is usually rarely used (only for the occasional peaks, that's the idea).

If range expansion is enabled then samples less than -3.11dBFS (<= 0x5980) are attenuated by 6dB, samples greater than -3.11dBFS are attentuation progressively less, giving the end mapping of 0dBFS to 0dBFS.

If range expansion is not enabled (including non hdcd's) all samples are simply attenuated by 6dB.

Then there's the reconstruction filter selection. The whole idea of this is at least questionable. (I seriously doubt that this constitutes any subjective improvement!)

I'm not sure how this works exactly. There are 3 bits left that could be used for controlling the reconstruction filters, although there were always zero in all my test cds, and wmp didn't seem to respond to them at all.

HDCD also supports some kind of dynamic range compression for the lower levels. It works like this: Quiet signals can be amplified in steps of 0.5 dB (0.0-7.5 dB = 4 bit side information) smoothly before quantization.

Exactly so.

The side informations (whether peak compression is used, which reconstruction filter to use and the compressor gain) is transmitted via the least significant bits with the help of a linear feedback shift register. The only information you can't find on the net is how they actually code this side information. So, you can't implement a proper decoder.

Just needed a little bit of cryptanalysis.

On a candidate hdcd, during the silent intro at the start, the left and right channels matched exactly with a delay of a few hundred samples. This delay was rather fortunate, as it was hence, possible to identify where an hdcd code had been inserted by looking for a section where the correlation between the two channels stopped briefly. The locations of these sections were such that it was clear that the hdcd codes had been inserted into the left and right channels at the same time, and it was hence possible to assemble a set of encoded hdcd codes. Guessing that hdcd codes would contain some sort of magic number to mark them, and that a maximal length LFSR would have been used, it was simply a matter of trying all maximal length LFSRs on the data and looking for some correlation between them.

Hence, being able to extract hdcd codes, it was then fairly straighforward to use wmp as a reference to determine exactly what each bit of the code did.

Chris