OT: Compression format shoot off

Rednroll wrote on 9/22/2003, 5:56 PM

I was interested in seeing what kind of frequency response that all the compression formats had. I did a study where I took Pink noise, which will give you a flat response across all frequencies. I then saved this .WAV to various compressed formats and compared the frequency responses. For those of you interested here's what I found. Sorry it might be a little difficult to read I did a copy and paste from MS Excel.

Format File Size Compression Comp ratio Frequency Response
.WAV 10.0 Mb Full Range
.PCA 8.79 Mb 1.4:1 Full Range
.MP3 2.29 Mb 320 Kbs 4.37:1 Full Range
.MP3 1.37 Mb 192 Kbs 7.3:1 Brickwall Low-pass at 20 Khz, slight roll off starting at 15Khz
.MP3 939 Kb 128 Kbs 10.65:1 Brickwall Low-pass at 15.8Khz
.MP3 470 Kb 64 Kbs 21.28:1 Brickwall Low-pass at 10Khz
.Ogg 972 Kb 128Kbs 10.29:1 Full Range with a +3dB High Freq boast starting at 10Khz
.Ogg 696 Kb 96Kbs 14.37:1 Brickwall Low-pass at 16.5 Khz with a +3dB 10Khz boast starting at 10Khz
.WMA 2.34 Mb 320Kbs 4.27:1 Brickwall Low-pass at 20.3Khz
.WMA 1.17 Mb 160Kbs 8.55:1 Brickwall Low-pass at 19Khz
.WMA 965 Kb 128Kbs 10.36:1 Brickwall Low-pass at 17.8Khz

From the data above in my opinion for compression to quality comparison, it seems like you get the best bang for your buck from the Vorbis .ogg format. I got a full range response at 128Kbs with a slight boast at the extreme high frequencies, but not bad for a 10:1 compression ratio. Now, I didn't do a stereo seperation comparison, that's another day. :-)

Comments

MJhig wrote on 9/22/2003, 7:03 PM

From my understanding *.pca is the only truly lossless compression format in the list meaning that the others compress by removing data, data that cannot be replaced once compressed then uncompressed so that upon each save > open > resave, further damage is done to all but *.pca.

I'm regularly seeing right at 50% compression with *.pca.

In a one shot scenario and then distribution, given my understanding of *.pca is correct it would not be the best choice but for archiving it seems to be the only choice if size matters (and the other team says it really does :-).

MJ

Rednroll wrote on 9/23/2003, 9:21 AM

I agree. I'm not sure how the .PCA compressed data works. I'm sure the compression is a lot better for something like music, especially if the compression is based off of repition of like samples. In any case I'm sure pink noise is the worst case scenario since it's neither repetitive, and has a full spectrum of frequencies at all times.

bgc wrote on 9/24/2003, 12:56 AM

Lots of lossless PCM compression algorithms use differential encoding (code the difference between audio samples) and run-length encoding for each actual sample (if it's a quiet sample, most of the bits are set to 0). This typically gets you in the 50% compression range. Random signals like pink noise randomly jump from sample to sample so you don't get lots of intersample similarity to take advatage of, and if the signal level is high, there's not a lot of zeros in each sample to run-length encode.
Compression schemes like MLP (DVD Audio) do stuff like this and throw in the kitchen sink to get 10% or more compression.