I've searched everything with SSRC, dithering, and changing bit-depth, but came up with nothing similar...

I took two files ripped straight from a CD & deglitched, and tried converting them down to 8-bit files. Here's the command line I used:

ssrc_hp --rate 44100 --bits 8 --dither 3 --pdf 1 0.7 <infile> <outfile>

After encountering the bug, I tried the normal ssrc.exe also, but it exists there too...

Anyway, here's what I get...

http://www.dressedforautumn.com/audio/SoaS16.png
This is the 16-bit resolution spectral analysis. Now, SSRC it.

http://www.dressedforautumn.com/audio/SoaS8.png
8-bit... Freaky. =/

Just to make sure it wasn't the song, I tried another song:

http://www.dressedforautumn.com/audio/BatM16.png
16-bit

http://www.dressedforautumn.com/audio/BatM8.png
8-bit

Has anyone else ever seen something like this? It only happens with the high-amp ATH noise shaping (dither type 3). Is this just a setting to be avoided, or is it a bug in SSRC?



Edit : removed the embedded pictures

The same here. It seems to be a problem of clipping during/after sample rate conversion. Amplify before applying SSRC by e.g. -1.5 dB and the problem will occur later (or not at all).

The SSRC dithering routines are very sensitive for clipping. AFAIK the code contains stuff to make sure there is no clipping during the conversion, but it seems it's not quite perfect.

The code in FB2K does not have this problem.

I haven't tested yet, but as tigre says, it could be clipping. Note that strong ATH noise shaping dither has several bits of amplitude (as opposed to just +-1 bit of flat dither), and in case of 8-bit audio (that just uses 256 levels) it can be equivalent to several dBs of noise added.

For 8 bit-audio, ATH-shaped (be it soft or strong) dither is not a good idea most of the times, it will add great amounts of high-pitched noise, that can fry you tweeters, apart from being annoying*. Use flat or triangular dither instead, the one that sounds best for you.

*Edit: well, the high freq. noise added is not so huge, and will possibly be not perceived. However, for long-term or high loudness listening your tweeters may suffer, I guess not if you play the music at very moderate levels.
*Edit: substituted the "no recommended" thing for "not good idea most of the times", it's up to you.

The four posted pictures were 1 MB (4 times 250 kB lossless png), that's too big for pictures inside a post. We generally try to keep them below 10 kB (=3 seconds for a modem user).

To save audiograms, use jpeg compression, and set the color scheme to 4:4:4 instead of 4:2:2, in order to avoid washing out the colors.

Examples : http://www.hydrogenaudio.org/forums/index....t=25#entry96121