Audibility of "typical" Digital Filters in a Hi-Fi Playback
Reply #728 – 2014-11-28 16:14:28
Good morning Arny .At 3.5 Khz and eyeballing it, the noise floor of the content is around -25 db spl and the peak 90 db. So the total range is 90+25=115. How did you get the 70 number? Nice job of cherry-picking from misleading data. Boy that is a grumpy way of starting a reply . I have now seen this phrasing used multiple times to dismiss out of hand various data point. It must be part of the parlance of this forum. It is another form of FUD so I hope we don't keep using the tactic when we are having an educated and deep technical discussion. More below.The noise spectral density is said to have been measured using a 1 Hz bandwidth, but in fact the ear hears in critical bands which are about 1000 times wider around 3.5 KHz. The well known fallacy of measuring the perceptual qualities of music or ambient noise in narrow, constant frequency bands is being exploited. Hmmm. Very, very odd comment Arny. How much a rock weighs has nothing to do with how much weight you can carry. The signal to noise ratio of a music track is a measurement. It has nothing to do with what part of it we can hear. Indeed you said it properly yourself: "The musical selection used in the Meridian tests was a good example - the noise floor was only about 70 dB below peak levels and thus easily handled with 16 bits and best practices." See? You made no reference to what we can hear. Of course you didn't read the graph right so let's look at a marked up version to make it clear: . Your comment regarding critical band is just voodoo psychoacoustics in this context triggered by not understanding the method by which we get like data. Stuart graphs have been normalized to make it possible to compare a peak tone to a noise level. They need no further conversion to be used for computation of the signal to noise ratio. Indeed the very reason for existence of that graph is to make this very point regarding test music track. Let's put all of that aside. We just went through JJ's slides that stipulated the ear's dynamic range at 114 db. So we need no further mistranslation of that. Our track has 115 db of dynamic range which means if we heard it live, we would be able to appreciate it all. A 24-bit recording system with real 20-bit response gives us that. A 16 bit system does not. And just to bring some variety into the discussion, here is another trusted source on this topic: From the Professor Vanderkooy's paper which we discussed earlier on his proposal to perform this test correctly:A Digital-Domain Listening Test for High-Resolution John Vanderkooy Department of Physics and Astronomy, University of Waterloo, Waterloo, Ontario, Canada N2L 3G1 Using non-subtractive flat triangular probability-function (TPDF) dither will add ?2/6 of noise power, so that the theoretical S/N ratio of 98 dB for 16-bit audio becomes about 93 dB. While this may be audible by some listeners at elevated listening levels in a very quiet environment, especially at the beginning of a musical selection or at its final decay, its significance for audio quality is not as critical as other aspects. This is referenced in Stuart's paper. As you recall from our previous discussions, it outlines how one would create a much more reliable test of high resolution against down converted 16/44.1. So I hope are done with the constant declarations that "music" is always 70 db or so and hence you don't need more than 12 to 13 bits. That is junk audio science if you forgive me for saying so.