Disadvantages of using large NFFT value in pwelch/spectrogram

I'm using the pwelch function to analyse power spectrums of an audio signal:

        w = 512;
        nfft = 4096;
        fs = 48.1e3;
        [px1,px2] = pwelch(data,w,round(0.7*w),nfft,fs);

I'm trying to understand the disadvantage (apart from computing time) of using large values of nfft, in online examples the window length is often used. I understand the length of nfft determines the frequency resolution used, and that this is to do with zero-padding the signal.

If someone could give insight into how this zero padding works (considering the window length is << nfft length), and how increasing NFFt (as computing time is not a constraint) could be a disadvantage that would be great!.

Answer 1

[Greg Dionne]

You can mean one of two things by frequency resolution:

1. The frequency spacing between computed estimates (sometimes called the "bin width")
2. The ability to resolve frequencies of the sampled signal independently.

When you zero-pad an FFT you essentially perform (periodic) interpolation between data points in the Fourier space. You can think of it as taking a long signal and convolving it (in the frequency domain) with the Fourier transform of the shorter window. If your window is rectangular, then you are essentially performing sin(x)/x interpolation in the frequency domain. You get a lot more estimates in the sense of #1, but they're very smoothed out.

You don't actually gain any frequency resolution in the sense of #2. You just get finer interpolation between points. The only way to gain resolution is to actually have more sample points of the original signal.