Yes, you are confused. Maybe one or a few of the other engineers will clue you in as to your error.
The fact remains that phase shift can be audible. Phase was shifted, the time domain waveform changed as it always must, and it sounds different in this case. In other cases, a shift in phase also changes the time domain waveform, but ear/brain system does not hear it as different. IOW, whether or not a phase shift is audible or not turns out to be time-domain waveform dependent. A general rule to the effect that it can never be audible is thus falsified.
The fact remains that phase shift can be audible. Phase was shifted, the time domain waveform changed as it always must, and it sounds different in this case. In other cases, a shift in phase also changes the time domain waveform, but ear/brain system does not hear it as different. IOW, whether or not a phase shift is audible or not turns out to be time-domain waveform dependent. A general rule to the effect that it can never be audible is thus falsified.
Last edited:
An FFT is more than a power spectrum. It has fully complex phase infromation. Throw away the phase information and yes an FFT isn't going to discriminate, but keep the phase and it absolutely will as it encodes the full input sample information, with mathematical rigour.
You should't malign Fourier transforms because some people and software hides the phases... Encourage them to plot phase information whenever possible!
@Mark Tillotson It was not intention to malign Fourier transforms. My comments about it were in relation to a warning given by Purifi:
Here is what they said:
I’d like to offer up a small demo to caution against this form of Popular Psychoacoustics. Below are two signals whose power spectra are completely identical. Only the phase relationships are different.
So my criticism is in relation to how people are popularly using and interpreting Fourier transforms. The common FFT distortion plots where harmonic spur magnitudes are analyzed as though they are independent of each other in terms of human perception are of questionable value. Even displaying phase of each one in a table does not help to give much useful insight into what the harmonic structure might sound like. A distortion crest factor number might help a little, not sure. Maybe best would be a time domain distortion residual?
EDIT: On the subject of the predictive value of FFTs for subject SQ, even if phase information is taken into account, I don't see AP claiming their machines measure SQ. From what I have read they tell the truth and say that the measurements provide "figures of merit" numbers. THD is one such figure or merit. So is SINAD. In this thread we have seen what Sean Olive had to say about those two things, and what Earl Geddes has said about them. Looks to me like there are problems that need solving in area of measurements and their interpretation. One person can't fix it alone. We see already in the past few posts what kind of reactions can be invoked. Also, Personally I don't like being accused in so many words of lying when I'm trying to explain something technical.
Here is what they said:
I’d like to offer up a small demo to caution against this form of Popular Psychoacoustics. Below are two signals whose power spectra are completely identical. Only the phase relationships are different.
So my criticism is in relation to how people are popularly using and interpreting Fourier transforms. The common FFT distortion plots where harmonic spur magnitudes are analyzed as though they are independent of each other in terms of human perception are of questionable value. Even displaying phase of each one in a table does not help to give much useful insight into what the harmonic structure might sound like. A distortion crest factor number might help a little, not sure. Maybe best would be a time domain distortion residual?
EDIT: On the subject of the predictive value of FFTs for subject SQ, even if phase information is taken into account, I don't see AP claiming their machines measure SQ. From what I have read they tell the truth and say that the measurements provide "figures of merit" numbers. THD is one such figure or merit. So is SINAD. In this thread we have seen what Sean Olive had to say about those two things, and what Earl Geddes has said about them. Looks to me like there are problems that need solving in area of measurements and their interpretation. One person can't fix it alone. We see already in the past few posts what kind of reactions can be invoked. Also, Personally I don't like being accused in so many words of lying when I'm trying to explain something technical.
Last edited:
In short the issue is a spectrum amplitide or power plot is not the whole story, however obtained (old-school frequency sweep, or digitize/FFT). And furthermore often there's no consideration/understanding for many when using FFTs of window function choice, nor that the noise floor cannot be measured without taking account of the bin size and noise-equivalent bandwidth of the window function.
For distortion as opposed to signal characterization, the residual plot in the time domain is powerful and not often seen, but it should be widely used as it gives more useful insights.
For distortion as opposed to signal characterization, the residual plot in the time domain is powerful and not often seen, but it should be widely used as it gives more useful insights.
Pretty much. I guess I would also like to see in chapter 1 of the book on it, the so-called motivational chapter, that there be some mention of old graphical ways of calculating DFTs. To me, it seems like the correlation method may provide the most intuitive insight into the process of transformation between time and frequency domains. I would like to see students do a problem or two with graphical DFTs to get a good, intuitive feel for how that works.
The two signals only differ in the phases of their spectral components, ie they have the exact same amplitude spectra. Changing the phase is of course changing the time domain waveform since the time domain and frequency domains are connected via the Fourier transform. It is correct to say that the two signals only differ in the phases of their spectra.
You don't need any fancy tests for this. If you couldn't hear phase differences, then you wouldn't be able to locate the source of a sound in 3d space.
Level differences and other cues can still provide localization capability when phase differences cannot be heard.
Mark, I have seen those. Thanks.
Human hearing evolved to help us survive. We use phase to determine direction, and we are good at it. That involves the same sound arriving at two different times to each ear. It is that simple. The articles you are referring to are flawed, period.
Human hearing evolved to help us survive. We use phase to determine direction, and we are good at it. That involves the same sound arriving at two different times to each ear. It is that simple. The articles you are referring to are flawed, period.
Okay, you reject Lars Risbo, and you also reject Scott Wurcer too?
Scott Wurcer:
You are right, changing the phase of the harmonics will change the envelope of the time domain signal. You can create some interesting effect with this type of manipulations, there might have been some on that AES flexidisk (don't recall). At the -100dB or less level I doubt these are audible.
https://www.diyaudio.com/community/...orch-preamplifier-part-ii.146693/post-3118421
------------------------------
Care must be taken in differentiating between flat small signal frequency response and time domain response. The crest factor of music can be changed by all pass networks.
https://www.diyaudio.com/community/...orch-preamplifier-part-ii.146693/post-2682948
So, isn't what Wurcer described exactly what is happening in the Purifi demo (for the amplitude modulation file only)?
Also IIRC correctly Scott referenced a couple of studies in one of the Blowtorch threads as he explained that phase shift was no longer believed to be inaudible at lower frequencies. Perhaps it involves the same mechanism as above.
Yes, that is correct and I don't claim otherwise.
However, Markw4 claims that two audio signals (not their spectra) differ only in phase and uses this as an example of our heiring sensitivity to audio signal phase.
Phase is a property defined in the frequency domain (spectrum) . We need to transform a time domain signal to the frequency domain to identify the phases. I don’t see a problem with Markw4’s formulation.
All pass filters change the envelope of eg speech signals. It was shown decades ago that insertion of all pass filters can be audible in some cases.
I may not know much about audio amplifier building, but I do know a fair bit about signals and systems.
As a side note,
Propagation delays that our ears use (not only the delay though) under the fourier transform become phase shifts. However this property holds "well" for bandwidth limited signals, ie continous ones. Try and see the spectrum of 10 cycles of a sine wave followed by silence, and then the obvious question becomes "wait, is 20Khz bandwidth enough?"
No, not really. A sine wave in the time domain can be easily define with a phase parameter. Phase DETECTION however, is a nonlinear process, (needs a multiplier, a switch or something else) and thus it is most easily derived from a fourier transformation.
Fourier is a real number to complex transformation. The plots you see show the magnitude only. So, yes and no? No, the fourier DOES contain phase information and thus it is reversible to get back to the time signal and it is 1-1 transformation. You can get ONLY the original signal. Yes, magnitude only plots is not the whole story.
As a side note,
Propagation delays that our ears use (not only the delay though) under the fourier transform become phase shifts. However this property holds "well" for bandwidth limited signals, ie continous ones. Try and see the spectrum of 10 cycles of a sine wave followed by silence, and then the obvious question becomes "wait, is 20Khz bandwidth enough?"
Markw4, I have zero interest in seeing who can win a debate here. Your entire argument doesn't have a lot to do with the purpose of this thread, and a great deal more with rehashing material from an earlier thread.
Why not consider the main purpose of the thread now. We've had our fun and I'm sure the OP isn't interested in arguments of how we determine distortion. He's looking at crossover distortion which is easily seen on an oscilloscope from the output of a THD analyser (which is also referenced in the thread title).
Why not consider the main purpose of the thread now. We've had our fun and I'm sure the OP isn't interested in arguments of how we determine distortion. He's looking at crossover distortion which is easily seen on an oscilloscope from the output of a THD analyser (which is also referenced in the thread title).
So, you don’t see amplitude modulation of one audio file and constant amplitude of other audio file, albeit with shifting phase (which does not produce any amplitude variation), as any difference? Mind you, we are not talking about their spectra difference, rather real audio files difference and reasons for their perception as different. To you and Markw4, they differ only in phase. Oh well.
To us and to Helmholtz, in his case going back to 1843. His "Ohm's Acoustic Law," includes the principle that: the quality of a tone depends solely on the number and relative strength of its partial simple tones, and not on their relative phases. The examples given by Purifi illustrate that there are exceptions to the so-called 'law.' That is what the recent discussion in this thread has been in relation to.
https://en.wikipedia.org/wiki/Ohm's_acoustic_law#:~:text=The law was proposed by,not on their relative phases.
Moreover, the entire focus of the recent discussion, including Power Spectra of signals, has been within the context of the formalism of the Fourier Transform. In that context phase is clearly property of frequency, not of time. Again, Helmholtz was talking about the phases of the partials in frequency domain. Same as for Lars and me.
Maybe that was not clear, and or maybe you were not familiar with the historical underpinnings of the subject.
With that said, I agree with anatech that its probably time to get back to the thread's main topic.
Last edited:
Hi Markw4,
Please remind me again of the subject of this thread? Is it "Fourier transforms and how to argue distortion measurements", or is it "THD measurements at the smallest output power where crossover occurs".
You are hijacking this thread. Please stop.
I get it, "dont use THD" is the argument. Its a good argument. However if THD=0 then it doesnt matter which metric we use though 🙂
