Audibility of "typical" Digital Filters in a Hi-Fi Playback
Reply #301 – 2014-11-18 15:11:57
Good morning there. I must say, that was the shortest goodbye I have seen .FFS, we each don't communicate in a vacuum. You've already been schooled by others on the matter. You say that as if it is supposed to be something negative. It better not be. I live for days I learn something from these forum interactions as opposed to dealing with male insecurities to get personal. You have some very knowledgeable members here who for sure could school me. Alas, as it happens, they are the ones without a copy of the paper so can’t speak properly to what the test is all about. As it turns out the people with a copy of the paper don't have expertise in this field so we are honestly reading gibberish from them. Not saying the above to put folks down. It is the reality of the work in front of us. It is not a school project like Meyer and Moran. Even though the paper describes a listening test, it is extremely dense with jargon of audio signal processing, statistics, and analysis. It is way beyond the knowledge level of a typical forum member who doesn't do this work professionally. I hope to remedy this situation below.Any test scores showing rectangular dither didn't affect the results? If you are asking this question, it means that there is still no understanding of what the paper was about some 12 pages and 300 posts later. So please allow me to give you the summary as to facilitate proper schooling, whichever direction it will go. Let's start at the top. This is the title of the paper:The audibility of typical digital audio filters in a high-fidelity playback system No, I am not being pedantic . Please pay attention to what I have highlighted: digital audio filters. That is the mission of the paper. They start with 192 Khz music samples, then apply two filters to it. One is to represent CD sampling rate of 44.1 Khz and the other, 48 Khz. Here is the paper itself in technicolor to make sure these key points are not lost:2.3. Signal processing and test conditions Two kinds of linear-phase FIR (fnite impulse re- sponse) filter were used, both of which operated at 192 kHz and both of which were implemented us- ing TPDF (triangular probability density function) dither at the 24th bit. For both FIR filters, the ripple depth over the passband was a maximum of 0.025 dB, and the stopband attenuation was 90 dB. The frequencies of the transition bands were 23500- 24000 Hz and 21591-22050 Hz, corresponding to the standard sample rates of 48 kHz and 44.1 kHz re- spectively. Fig. 2 shows the amplitude and energy of the impulse response for the 48-kHz filter. I have highlighted two key things: 1. The "ideal" TPDF filtering was used contrary to the impression left so far in this thread but I will get to where it is not used and source of confusion in a bit. 2. The filtering is the kind of textbook response that we always say is as good as “perfect” and hence inaudible: a fraction of irregularity over the audio band and strong rejection of out of band spectra. This is what the test sets out to find out. Remember the title again: it is all about digital audio *filtering*. Key thing to note is that the above transformations are in 24 bit as is the source. Further, the sampling rate in all cases remains at 192 Khz. We are simply looking at what would happen if the bits were filtered down to lower bandwidth that 44.1 and 48 Khz would entail while keeping everything else the same. Our camp’s position is that no double blind test would ever demonstrate anything an audible difference to statistical confidence. As otherwise it would say the mere conversion of higher bandwidth to CD’s (and that of 48 Khz) is a lossy audible conversion. Let’s see how it turned out. The paper explains six (6) ways this filtering was tested: Translating into English, listing tests 1 and 4 (“none”) examine exactly what I described above. Filtering of a 192 Khz/24-bit file down to what it would have been at 44.1 Khz and 48 Khz sampling while remaining in 24-bit mode. Again please allow me to remind you that in both cases TPDF dither is used, not rectangular. So the answer you ask in this regard is “no.” RPDF was not used so it did not interfere with this part of the test. The other test conditions deal with what would happen if we converted the filtered 24 bit output to 16 bit samples using two different methods. One by simply truncating it or the so called “self-dither.” The extra resolution bits are simply discarded. This is tests #2 and #5. The other method and what folks have been fixated on, namely conversion of 24 bit to 16 rectangular PDF dither. The theory says this is superior to doing nothing. Results here are surprising as with the filtering. I hope everyone is with me so far. That we have six (6) independent tests, not just one like Meyer and Moran performed. This allows us to investigate the effect of each processing separately and together.The Results Here they are: Remember, we had three tests for each filtering: no dither (24 bit), truncation to 16 bits, and dithered to 16 bits. And that is what is represented in the above graph. The horizontal line at 56% shows the 95% level high confidence bar. Please don’t be swayed by the small numerical value of “56%.” The statistics are not for lay people. Just remember that this is the same standard we require of all such tests. If we achieve better than 95% results, we trust the outcome. We see that five (5) out of six (6) listening tests comfortably cleared the 95% interval both in their mean and standard deviation. Oddly, the 24-bit, 48 Khz sampling did not do so fully when you account for the error range. A further look at the results focusing on critical/more revealing music segments dealt with that by bringing that score above 95%. So what do we learn? All processing was distinguished from the original to statistics confidence! Filtering with or without conversion to 16 was audibly different to listeners (to statistical confidence). None of this was expected by our camp’s vocal members. And certainly counter to Meyer and Moran’s test indicating these are all transparent conversions. The rest of the paper theorizes as to why these differences were heard. I won’t get into that now. But I hope it is clear that this test easily schools all of us. That our knowledge of psychoacoustics with respect to digital audio filters/quantization is not complete. All of these tests should have failed to invalidate the null hypothesis but they did not. I am confident this is one of the reasons that peer reviewers were impressed with this work. It is groundbreaking in documenting audible differences this way. Circling back to your question, if you like to throw out the results of 16 bit conversion with or without dither, do. But then explain why simple filtering of spectrum to 22.05 and 24 Khz resulted in audible differences despite near perfect frequency domain response of this type of anti-aliasing filter.