Applied DSP No. 3: Short-Time Fourier Transform

[@bigmistqke]

These videos are really great. Not a lot of approachable and clear DSP content around, but these ones really hit the sweet spot!

[@arullin]

Clarity with visualisation....superb presentation

[@phil.s3713]

This video and the editing are amazing! Thank you for taking the time to make it.

[@bradlanducci9011]

As a CS major who is learning this stuff for audio programming, I find these videos very understandable. I have no real background in DSP, but I've been watching your videos and others like it and these stand out by far as the most informative.. thanks!

[@fano72]

High quality educational content!

[@oscarwjy5084]

OMG finally I understand the concept of STFT and spectrogram, this really help a lot for my final year speech enhancement project tqvm!

[@AlexGarcia-cr5st]

wow. I loved your intro! Amazing job

[@yesidcanoc]

Thank you for this great lecture.

[@netanelmad]

Great videos! Thank you so much

[@안준표-o5p]

좋은 강의 잘들었습니다. 감사합니다.

[@irelandtamilandamo]

Really a great lecture !!

[@antoniodiaz3896]

👏👏👏 bravo!!!

[@nayanvats3424]

Are we going to discuss about Single Frequency Filtering, which uses a filter bank approach to overcome the shortcoming of STFT and can provide better frequency and time resolution ?

[@TheGmr140]

wonderful video

[@eduardojreis]

Great series! Much much appreciated! I got a question: 09:07 - It is said that a frequency could occur at anywhere, implying there are segments that frequency does not occur. However, my current understanding is that the FFT of the full signal assumes each frequency component is present at ALL times (not SOME timer, nor anytime), with its particular magnitude. Right?

[@austin_ma]

Thank you sir!!!!

[@71sephiroth]

wish I could apply for this course with all the materials....

[@세오y]

amazing, lecture.....shout it out!

[@ytubeleo]

Excellent video. At 07:00, instead of padding with zeros, could resolution be improved by just repeating the original signal section several times and taking the Fourier transform of that? Would that increase our resolution or not work? For example, if I have a 10 second ECG recording, would it be sensible to repeat it 10 times and do the Fourier transform of that to try to get better resolution? For example, when trying to find the fundamental frequency more accurately. What would you recommend? Would it work to stretch out the signal - interpolating it - for example to ten times the number of samples (or increase the sample rate) to get higher resolution? Apart from saving on computation of interpolation, does zero-padding give more accurate estimations of peaks? Thanks!

[@spasmodia]

If I'm not mistaken, zero padding is the process of interpolation and youre essentially adding "zero data" in-between samples to bridge these samples. Repeating the same signal will just give you the same Frequency Response which you've already gotten from the first sample.

[@youngmoo-kim]

No, repeating the period would give you more frequency samples, but it wouldn't increase actual frequency resolution because the Fourier Transform is already the infinite periodic replication of your signal. Similarly, the zero padding (extension of the signal with zeros to create a larger "base period") doesn't add new information, but just increases the samples of the Fourier transform to reveal what's already there. For fundamental frequency estimation, the best approach is to take whatever signal you have (10 sec) and then zero-pad with enough zeros to get a better resolution of the frequency peaks. Interpolation/resampling of the signal would not provide higher resolution... the frequency content (and total information) of your signal is fixed. No amount of DSP will add useful information, so the goal is to most clearly reveal the info that's already there. Hope this is helpful!

[@youngmoo-kim]

@@spasmodia Just to be clear, zero-padding is adding zeros to the end of your signal "frame" (not in between the samples of the signal itself... that's a form of upsampling). You're correct that the frequency content when upsampling remains the same, it just changes the position of the aliases. This is, in fact, part of the process I used to create the aliased sound samples.

[@classicalharmonicanalysis3348]

Small typo at 2:15 - instead of "t" it should be "n" and get rid of "dt". Otherwise neat video :)