The Sliding DFT generates the spectrum for every input sample. Using the FFT, the spectrum is generated only after a certain number of samples (N) are obtained. It appears to me as though the cycles to perform the Sliding DFT would be much larger than the cycles required to do an FFT, to generate an N point DFT. Would it be recommended to use the Sliding DFT to generate an N point spectrum? The motivation for me to consider using the Sliding DFT to generate an N point spectrum is time resolution.

Could anyone give me an idea of the challenges in implementing the Sliding DFT on a DSP?

  • $\begingroup$ For Audio? Storing an entire N point spectrum per audio sample results in a lot of data, so streaming the DFT results to/from memory would be a major constraining factor. Also, more overlap can be used with the more efficient FFT to gain some time resolution. $\endgroup$
    – hotpaw2
    Feb 24 '16 at 16:25
  • $\begingroup$ my only concern with the sliding DFT is a similar concern for the moving average filter implemented the efficient way, with an accumulator. if you cannot guarantee that what falls off the edge of the delay line (and is subtracted from the accumulator) is not exactly what was added before, including roundoff error, you might get stuck with little turds in your accumulator that will never come out nor will die away. $\endgroup$ Feb 25 '16 at 0:54

The paper Sliding DFT control algorithm for three-phase active power filter implements a sliding a sliding DFT on a TMS320F2812 from Texas Instruments, and discusses some advantages, and An Enhanced Interpolated-Modulated Sliding DFT for High Reporting Rate PMUs does similar work on FPGA.

From what I remember, you may have a look at the stability, see The Sliding DFT.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.