OK, let's do the numbers. Assume inter-ear distance is 8". Distance from R speaker to R ear is about 118.1". Ditto L speaker to L ear. Move the head an inch. Now the distance from R speaker to R ear is 118.6", L speaker to L ear is 117.6".
How about the cross terms? Centered, the L speaker to R ear is 122.1", as is R speaker to L ear. Move that inch, and L speaker to R ear is 121.6", R speaker to L ear is 122.6".
Attachments
I said don't need no stinkin' FFT's the final waveform was produced without them, simple adding of waveforms. The FFT length does not matter the spectral amplitude distribution obeys that expected by noise, short FFT's more variance. These were 64K with Auditions rolling average used to see a clear floor. Since I operated on a single tone the averaging does not obscure the harmonics if present.
So you mentioned terrible fuzzy distortion, suggest a test. Triangles and square waves don't count. Southpaw's harmonica would be great IMHO.
Last edited:
That is what we should work on. It may just be possible there are more productive uses for your time...
One last comment, when the -80dB sine wave is subtracted from the "horrible" waveform and my final waveform the total rms power is identicle (within .5dB), So there is no free lunch, the floor has been turned into noise instead of being concentrated into tones at the theroretical 16 bit level. So the dither seems to work even on a sample by sample basis and Barry was right.
I did. You can derive that from the numbers I gave. You can even consider pan-potting and the effects of the head shift on where the image will be perceived, just as you can with the simple time delay in one channel (that doesn't actually happen in any but the cheapest and oldest DACs). In fact, since the intensity shifts here are negligible, the small head shift reduces to the simple time delay.
Centered left ear to left speaker 118.118" to right speaker 121.115"
Centered right ear to left speaker 121.115" to right speaker 118.118"
Agreed.
Offset 1" Now left ear to left speaker 117.648 to right speaker 122.642
Offset 1" Now right ear to left speaker is 122.642 to right speaker 117.648
Agreed
Now what does each ear perceive? Obviously it hears the combination of sound in terms of time, amplitude and frequency. As is often modeled we are concerned with the lower frequencies as there is ample evidence that at higher frequencies other localization factors come into play.
So centered both ears are averaged at 119.617" of time delay. Offset one inch moves the average to 120.145" But here is the kicker, the ears are not comparing that movement, because they one isn't hearing the first number and the other the second. They are both still hearing the same combined number. So they don't detect the change.
So how far do they have to move so that one ear gets a .1356" difference in the path length average from the two loudspeakers?
SY,
I thought the question is why don't things sound different when you move your head fractions of an inch. Which was logical based on the 10 us perception claim.
My point was that Pythagoras guy answered the question even before Curl was born.
I understand that sometimes what is clear to me is not being communicated by me to others clearly.
Now if you want to know what happens when two sound sources have a 10 us shift between them, you start getting a notch at some above band frequency so that probably is not a good test as there are valid arguments as to what is the upper limit of hearing or is it detecting phase shifts?
It might be worthwhile to produce stereo tones with harmonics offset differently.
I can tell you from experience I can set an audio delay in a multiple loudspeaker setup to better than 3 ms by ear. Yes that is a great deal worse than 10 us.
So lets do the college test. I'll blindfold you and throw water balloons at you. But to be fair, they will have a noise source on them. We can see what resolution you require to avoid getting wet.
Well it is time to quit. Have fun.
No, the question was, if there's an interchannel time delay of 10us in a stereo setup, can that be perceived with speakers in a room? Pavel suggested darkly that this is a problem in DAC, but wouldn't elaborate. The speaker/head-positioning was a way of showing the scale of what's actually a non-problem- the problem reduces to being rather identical with speaker and head positioning.
The question which was not asked explicitly, but I've answered anyway, is do you get 10us interchannel delays from ordinary electronics? And that's clearly "no."
edit: I've done a recent DBT for phase shift/delay. I was able to hear it, so was Pano, but we were highly amused that the Golden Ears refused to even try.
The question which was not asked explicitly, but I've answered anyway, is do you get 10us interchannel delays from ordinary electronics? And that's clearly "no."
edit: I've done a recent DBT for phase shift/delay. I was able to hear it, so was Pano, but we were highly amused that the Golden Ears refused to even try.
Scott,
I believe you want to look up "subtractive dither"
it would have been great if the RedBook standard's committee had thought to define metadata, sync frame marking for subtractive dither - the horsepower was beyond conceptions consumer level digital at the time - today it would be no problem – it also helps to actually have better than 16 bit ADC, DAC at both ends of the process
the Wannamaker 1997 PhD thesis seems to be a fairly useful source – I think I can follow about ½ of the math
Howard,
I’m pretty sure I’ve got the definition of TPDF dither amplitude correct in http://www.diyaudio.com/forums/anal...ch-preamplifier-part-ii-1732.html#post2766594
I’ve checked with SciLab (free MatLab workalike)
a +/-1 count TPDF dithered –60 dB sine gives 33.3 dB S/N (where noise is the residual from subtracting the unquantized –60 dB sine)
for a 93 dB S/N
as a check the same method without dither gives 97.95 dB which agrees well for a finite sim, random number generator calc compared with the “official” equation
16* 6.02 + 1.76 = 98.1 dB
most of this discussion would be pretty pointless if even 2x higher sample rate, 24 bit DVD-A were common today as a consumer distribution format
but listening tests so far don’t show CD audio as “night and day” audibly different from higher res - not like good vinyl, tape, CD “hifi” sources are improvements over Edison wax cylinders or 300-5kHz 8-bit companded telephone voice BW
I believe you want to look up "subtractive dither"
it would have been great if the RedBook standard's committee had thought to define metadata, sync frame marking for subtractive dither - the horsepower was beyond conceptions consumer level digital at the time - today it would be no problem – it also helps to actually have better than 16 bit ADC, DAC at both ends of the process
the Wannamaker 1997 PhD thesis seems to be a fairly useful source – I think I can follow about ½ of the math
Howard,
I’m pretty sure I’ve got the definition of TPDF dither amplitude correct in http://www.diyaudio.com/forums/anal...ch-preamplifier-part-ii-1732.html#post2766594
I’ve checked with SciLab (free MatLab workalike)
a +/-1 count TPDF dithered –60 dB sine gives 33.3 dB S/N (where noise is the residual from subtracting the unquantized –60 dB sine)
for a 93 dB S/N
as a check the same method without dither gives 97.95 dB which agrees well for a finite sim, random number generator calc compared with the “official” equation
16* 6.02 + 1.76 = 98.1 dB
most of this discussion would be pretty pointless if even 2x higher sample rate, 24 bit DVD-A were common today as a consumer distribution format
but listening tests so far don’t show CD audio as “night and day” audibly different from higher res - not like good vinyl, tape, CD “hifi” sources are improvements over Edison wax cylinders or 300-5kHz 8-bit companded telephone voice BW
Last edited:
I lied, just to explain those little spurs on the high end. That is probably the gaussian vs TPDF and the -90dB is not the exact amount. On TPDF, apriory you have no way of telling noise in amplitude or time so TPDF which is the convolution of two uniform distributions applied only to the amplitude creates a uniformly distributed vector error at the sample time that way the quantization error is truely spread over the whole spectrum rather than concentrated at tones. Using algorithmic generation of the noise with a seed you basicly have an infinitely compressed noise signal so information theory is not violated.
jcx - I don't read a lot of audio journals so I was playing with that idea from 1978. I'm sure the theroy has been expanded. I got to .5dB theroretical playing with Audition OK by me. Do you like my last arguement? I was trying to make an intuitve connection to how this all amounts to a simple additive noise in two dimensions.
Last edited:
Trying to put the topic behind me, I guess the piece of information I am missing here is: how much dither is necessary to remove quantizing error as a function of bit depth.
I can't find a definite answer in any of my references, so if someone could shoot me a quick answer to this, I'll stop ragging on the subject
If this was answered yet I missed it. One answer is that if the dither could be made with a rectangular probability density function an RMS level of 1/3 LSB would be exactly enough to remove all quantization artifacts. Other PDF's will require more (I think! - there's a big world out there and much escapes me). One practical PDF works with a peak to peak level of 2 LSB.
Somebody who knows better is bound to speak up with corrections and details. Dither has been around since about 1950 (in video) but they didn't (overtly) use any in 1970 in the US Army PCM telephone systems I worked on.
All good fortune,
Chris
edit: answer appeared while I was typing!
Last edited:
- Status
- Not open for further replies.