Understanding FFT in Audio Measurements

[amirm]

FFTs are the fundamental to many measurements I show in my reviews. You can search for it online but I am not sure you will find any easy to understand explanation of it that applies to the way we use them in audio measurements. So I thought I do a walk-through using the Audio Precision of what it means, and how its parameters can drastically change what it displays.

As I asked the video, please let me know if you like these kinds of technical tutorials. To be sure they are not as much "fun" to watch as audio debunking ones. So like to know if you want to see more of them or not. And whether you could follow this one which is probably one of the most difficult concepts we have in audio measurements.

While I have not done a full text write up on this, the concept is included in my tutorial on digital audio measurements: https://www.audiosciencereview.com/...derstanding-digital-audio-measurements.10523/

 

[BigSweeny]

As I asked the video, please let me know if you like these kinds of technical tutorials. To be sure they are not as much "fun" to watch as audio debunking ones. So like to know if you want to see more of them or not.

Are you asking people on audiosciencereview.com if we want to see more technical information? No, we'd prefer more subjective opinions about how "musical" the measurements sound Seriously though, these videos are great @amirm. As you already said, people can read this information online but it's nice to have it in video form from someone who has the experience that you have.

With regards to what kind of content your videos should contain; please do whatever you want to do.
The reviews are an excellent source for mentioning the basics of what you're measuring ("SINAD below 115 is below human hearing, this DAC has 120 so it's fantastic") to give the audience an understanding of is X product worth the money.
The technical videos are nice because past the basics it's good to understand how SINAD/FFT/X measurement is calculated and what it means so one can discuss it with others or follow along at a more 'intermediate' level.
The debunking videos are good for simply answering the question of "will I get an audio improvement if I get X audiophile thing" and they're also funny to see how much misinformation/nonsense there is in the audio world.

If you ever feel the desire to go further indepth, such as having online meetings with engineers from companies that you respect and talk technicals of insides of products, why they made X decision or what improvements they are attempting, then that would be amazing as well.

 

[diddley]

Well done video, i think it is even better then debunking video's.

 

[JustAnandaDourEyedDude]

This video lays fundamental groundwork to help audiophiles and those of the general public that are interested in the performance of their audio gear to understand frequency domain measurements, without needing a STEM background. Kudos on an exemplary explanation that discusses the essentials and illustrates the key measurement parameters without throwing any math at the viewer. The flow of ideas is logical and sounds unscripted and natural, and the quality of presentation is on par with Monty Montgomery's justly famous video on digital representation of music. Very compelling point 23 minutes in when you explain that the outcome of the Fourier decomposition into sine waves together with the window functions and the sampling parameters makes the frequency domain analysis applicable even to square waves and to complex music in general, and relatable to psycho-acoustic measurements.

Until now, the interpretation of your measurements graphs was likely still a little beyond the average reader's capabilities, and they largely had to rely on your stated conclusions in each review. I think this video begins to enable the average reader to feel secure in understanding for themselves what your measurements show about any particular audio component. More videos like this one that explain the rest of the basic technical details in audio science and audio gear measurement and their significance in lay terms will be welcome and important in giving the lay reader/viewer confidence in the rationality, reliability and consistency of the science-based and measurement-based approach to evaluating audio equipment.

 

[beefkabob]

This is probably the first video I thought was actually interesting!

 

[Hayabusa]

Hi Amir, nice video!

Talking about FFTs, I always wondered if some measurements done by long FFTs (1M samples for instance) could hide audible artifacts in signal.

To mention an extreme case:
Lets say we have 1M samples of a 1 KHz sine wave and we add one FS dirac pulse when the sine wave is -FS.
Clearly this audible as a sharp click.
In the 1M FFT spectrum though it will show up as a -120dB white noise floor.
Never did the experiment, but my back of the evelope calculations are right isn't it?

 

[Blumlein 88]

Hi Amir, nice video!

Talking about FFTs, I always wondered if some measurements done by long FFTs (1M samples for instance) could hide audible artifacts in signal.

To mention an extreme case:
Lets say we have 1M samples of a 1 KHz sine wave and we add one FS dirac pulse when the sine wave is -FS.
Clearly this audible as a sharp click.
In the 1M FFT spectrum though it will show up as a -120dB white noise floor.
Never did the experiment, but my back of the evelope calculations are right isn't it?

Yes, long FFT's can hide momentary events.

That is why using Spectrograms, which are running color coded FFT's of the signal need to have FFT size chosen with care. Usually shorter is better for seeing non-steady signals in a spectrogram. Too long an FFT can miss it completely.

At the top is a 1k FFT of a few seconds of music. Bottom is the same only with a 32k FFT. You can see on the right half how regular percussive sounds with a 1k FFT have been smoothed out and nearly disappear in the 32 K FFT.

 

[NYfan2]

Thanks Amir, I really enjoyed this video.
Although I think the debunking videos are also great because these show the public that most 'audiophile tweaks' are BS.

I only wonder if non-technical viewers understand what is happening, I am an electrical engineer so I know the basics of sampling frequencies and sine waves and block waves etc. I am not sure if non-technical viewers can keep up, it will be interesting to hear that.

 

[BDWoody]

Great choice for a video. It's like you are pulling back the curtain to demystify what for many is like a foreign language.

Educating the consumer... Love it.

 

[Tks]

Yeah I really liked this video. Especially with newer people getting into audio, I had one person tell me "I have no idea what I'm looking at" when I referred them to the site..

Speaking of all the versatility about FFT conversions and insights. Would this allow us to do something like figuring out what what inherent distortions of say a headphone, are doing to the overall frequency response of the headphone? Like if you had a way of representing an FFT of a device with distortions/noise arifacts removed, what happens to the overall measured frequency response?

Is that even possible with current hardware/software of the AP?

 

[Hayabusa]

Yes, long FFT's can hide momentary events.

That is why using Spectrograms, which are running color coded FFT's of the signal need to have FFT size chosen with care. Usually shorter is better for seeing non-steady signals in a spectrogram. Too long an FFT can miss it completely.

At the top is a 1k FFT of a few seconds of music. Bottom is the same only with a 32k FFT. You can see on the right half how regular percussive sounds with a 1k FFT have been smoothed out and nearly disappear in the 32 K FFT.
View attachment 115416

But this means that if a device is measured by the AP that has for instance a DSP bug in its firmware that sends out a wrong sample every few seconds or so, it would measure perfectly well and still have audible artifacts in its output. So only based on this you can not claim transparency. Some how there has to be an extra test that calculates the peak error between the reference and measured signal in time domain to capture these events.

 

[vkvedam]

Excellent!

 

[beefkabob]

Measurements don't necessarily tell you when something is broken. A Ferrari can do 0-60 in 3 seconds while belching black smoke out the back. It still gets a bad review.

 

[Ajax]

Hi Amir,

Any information that you can provide that educates consumers in how to make better choices is welcome. I would therefore encourage you to provide a video analysis of each of your graphs and put them all together in a "technical" folder for easy access.

Knowledge and transparency is real power in any of life's endeavours.

 

[Hayabusa]

Measurements don't necessarily tell you when something is broken. A Ferrari can do 0-60 in 3 seconds while belching black smoke out the back. It still gets a bad review.

But in this case these kind of errors can be measured by the sampling method like the AP. If only there would be an extra feature called 'peak error time domain" we can at least detect these kind of DSP firmware errors.

 

[MKreroo]

I was thinking about suggesting these contents a small while ago, nice to see that we have these now.
I think it will be interesting if we can get videos talking about electronic and circuitry too

 

[fieldcar]

One thing that I've always wondered is that during your multi-tone test, do the tones mask potential harmonic THD, or are the tones offset at non-harmonic increments? EDIT: Nevermind. It looks to be offset by a small amount.

https://www.audiosciencereview.com/...ments-line-out-multitone-balanced-png.114520/

 

[phoenixsong]

FFTs are the fundamental to many measurements I show in my reviews. You can search for it online but I am not sure you will find any easy to understand explanation of it that applies to the way we use them in audio measurements. So I thought I do a walk-through using the Audio Precision of what it means, and how its parameters can drastically change what it displays.

As I asked the video, please let me know if you like these kinds of technical tutorials. To be sure they are not as much "fun" to watch as audio debunking ones. So like to know if you want to see more of them or not. And whether you could follow this one which is probably one of the most difficult concepts we have in audio measurements.

While I have not done a full text write up on this, the concept is included in my tutorial on digital audio measurements: https://www.audiosciencereview.com/...derstanding-digital-audio-measurements.10523/

Wow I'll never see graphs the same way- it's like they've come alive

 

[Lambda]

Importend and good video!

@amirm
Why not use (dB)V/√Hz so the measurement is directly comparable without knowing fft gain
https://www.audiosciencereview.com/forum/index.php?threads/a-better-way-to-graph-noise-fft-gain-db-v-√hz-1-3-octave-band-noise.20702/

To mention an extreme case:
Lets say we have 1M samples of a 1 KHz sine wave and we add one FS dirac pulse when the sine wave is -FS.
Clearly this audible as a sharp click.
In the 1M FFT spectrum though it will show up as a -120dB white noise floor.
Never did the experiment, but my back of the evelope calculations are right isn't it?

Your right. i have asked the same question here
There must be a better way to do peak detection.

Measurements don't necessarily tell you when something is broken. A Ferrari can do 0-60 in 3 seconds while belching black smoke out the back. It still gets a bad review.

Non periodic noise is not necessarily a "defect" but something that's not showing up in the FFT.

 

[Dj7675]

Great video. These kinds of videos are really important to readers with no technical background (like myself). As you plan out what videos to do, I think audibility will be some of the most important content. One article to put in video form might be:
https://www.audiosciencereview.com/forum/index.php?threads/dynamic-range-how-quiet-is-quiet.14/
One of the biggest things often stated is my room noise floor is “x” so all modern devices are transparent...Getting to the point of why measurements matter and your article on dynamic range I think helps explain that really well and would be even better in video form for the average user.