16 bit vs 24 bit
Reply #69 – 2008-11-27 04:27:07
Voxengo's software says that 8-, 16- and 24 bit output is dithered with a gaussian noise and a slight noise-shaping. I have no idea if it is the proper way to do it. The shape of the dither is quite significant. I'm not sure a "slight" noise-shaping is sufficient. I respect the difficulty in obtaining these results. However, I am afraid that all they show is that this dithering algorithm is discernible from the high-resolution original. If I were to perform this test, I'd try it with foobar2000's dither. Nonetheless, thank you for the results. As I've always considered 24/96 to be excessive, you've given me a point to ponder. I guess your listening conditions are pretty extreme and uncomfortable in terms of overall loudness, Pio, but roughly as loud as you'd need to have any chance of ABXing 16-bit from 24-bit. i.e. you need to have the programme peaks about as loud as a petrol (gasoline) chainsaw at arm's length without ear defenders (113-116 dBa is typical) to be able to come close to picking up quantization, dither noise or noise modulation at the LSB level without noise shaping while you listen to the fade-in. With optimal dither and strong ATH noise shaping, you probably need to near the pain threshold. Gaussian pdf dither isn't optimal. Triangular dither -1 to +1 peak to peak (sum of two independent rectangular dithers of -0.5 to +0.5 peak to peak) is optimal for the first and second moments, which prevents quantization distortions and noise modulation - the only two effects (moments) of bad dither that are considered audible. Gaussian might be more audible (higher energy) while failing to completely eliminate both effects. I'd agree that foobar2000's dither options, esp on older versions, should allow a better test of theoretically valid dither. Versions around 0.8 allowed:no dither (i.e. rounding/truncation quantization) flat dither (no noise shaping) - i.e. triangular pdf, dither with soft ATH noise shaping (less HF boost), dither with strong ATH noise shaping (recommended setting in fb2k). I think the range of dither/shaping options has been reduced and simplified in the 0.9 versions of fb2k. From calculations based on commonly-accepted figures for human audition (difficult for me to test in practice with my equipment and noise environment) I'd expect flat dither (triangular pdf) to be inaudible in comfortable listening but near the borderline of audibility/ABXability at extreme but bearable loudness. I'd expect good valid dither with noise shaping (not noise shaping alone) to be practically inaudible unless you selectively turn up the fade-in to extreme levels and turn it down to avoid painfully loud peaks during the main programme material. This would probably remain the case for tracks with replay gain values as positive as +12 dB or even a little greater (i.e. 77 dB SPL calibrated, or 6 dB below the 83 dB SPL reference pink noise specified in the Replay Gain calibration spec and used in movie theatre loudness calibration). Such quietly-recorded tracks could typically be increased by about 12 to 15 dB without clipping, but their noise-shaped dither is probably about 15 dB or so below the level of spectrally-flat (triangular pdf) dither.I'd be very interested to see disconfirming evidence (ABX) from you or anyone else and to know how the results vary with different dither and noise shaping options.