To obtain insight into the operation of filter banks implemented using an FFT, this section will derive the details of the DFT Filter Bank. More general STFT filter banks are obtained by using different windows and hop sizes, but otherwise are no different from the basic DFT filter bank.
where is the input signal at time , and . In this section, we will show how the DFT can be computed exactly from a bank of FIR bandpass filters, where each bandpass filter is implemented as a demodulator followed by a lowpass filter. We will then find that the inverse DFT is computed by remodulating and summing the output of this filter bank. In this way, the DFT filter bank is shown to be a perfect-reconstruction filter bank. The STFT is then an extension of the DFT filter bank to include non-rectangular analysis windows and a downsampling factor .
In this form, it is clear why the filter (9.5) is called ``running sum'' filter. Dividing it by , it becomes a ``moving average'' filter, averaging the most recent input samples.
so that its frequency response is
Recall that the term is a linear phase term corresponding to a delay of samples (half of the FIR filter order). This arises because we defined the running-sum lowpass filter as a causal, linear phase filter.
We encountered the ``aliased sinc function''
previously in Chapter 5 (§3.1.2) and elsewhere as the Fourier transform (DTFT) of a sampled rectangular pulse (or rectangular window).
Note that the dc gain of the length running sum filter is . We could use a moving average instead of a running sum ( ) to obtain unity dc gain.
Figure 9.11 shows the amplitude response of the running-sum lowpass filter for length . The gain at dc is , and nulls occur at and . These nulls occur at the sinusoidal frequencies having respectively one and two periods under the 5-sample ``rectangular window''. (Three periods would need at least samples, so doesn't ``fit''.) Since the pass-band about dc is not flat, it is better to call this a ``dc-pass filter'' rather than a ``lowpass filter.'' We could also call it a dc sampling filter.10.1
Modulation by a Complex Sinusoid
Given a signal expressed as a sum of sinusoids,
then the demodulation produces
We see that frequency is down-shifted to . In particular, frequency (the ``center frequency'') is down-shifted to dc.
Figure 9.13 shows how a bandpass filter can be made using a lowpass filter together with modulation. The input spectrum is frequency-shifted by , lowpass filtered, then frequency-shifted by , thereby creating a bandpass filter centered at frequency . From our experience with rectangular-window transforms (Fig.9.11 being one example), we can say that the bandpass-filter bandwidth is equal to the main-lobe width of the aliased sinc function, or radians per sample (measured from zero-crossing to zero-crossing).
Since the bandwidths, as defined, are , the filter pass-bands overlap by 50%. A superposition of the bandpass frequency responses for is shown in Fig.9.14. Also shown is the frequency-response sum, which we will show to be exactly constant and equal to . This gives our filter bank the perfect reconstruction property. We can simply add the outputs of the filters in the filter bank to recreate our input signal exactly. This is the source of the name Filter-Bank Summation (FBS).
Comparing this to (9.12), we see that the filter-bank output , , is precisely the DFT of the input signal when , i.e.,
In other words, the filter-bank output at time (the set of samples for ), equals the DFT of the first samples of ( , ). That is, taking a snapshot of all filter-bank channels at time yields the DFT of the input data from time 0 through .
When for any integer , the Sliding DFT coincides with the DFT filter bank. At other times, they differ by a linear phase term. (Exercise: find the linear phase term.) The Sliding DFT redefines the time origin every sampling period (each modulation term within the DFT starts at time 0 for each transform), while the DFT Filter Bank does not redefine the time origin (modulation terms are ``free running'' as they would be in an analog filter bank). Since ``DFT time'' repeats every samples, the two treatments coincide every samples (i.e., for every integer ).
When is a power of 2, the DFT can be implemented using a Cooley-Tukey Fast Fourier Transform (FFT) using only operations per transform. By keeping track of the linear phase term (an modification), a DFT Filter Bank can be implemented efficiently using an FFT. Uniform FIR filter banks are very often implemented in practice using FFT software such as fftw.
Note that the channel bandwidths are narrow compared with half the sampling rate (especially for large ), so that the filter bank output signals are oversampled, in general. We will later look at downsampling the channel signals to obtain a ``hopping FFT'' filter bank. ``Sliding'' and ``hopping'' FFTs are special cases of the discrete-time Short Time Fourier Transform (STFT). The STFT normally also uses a window function other than the rectangular window used in this development (the running-sum lowpass filter).
This suggests that the DFT Filter Bank can be inverted by simply remodulating the baseband filter-bank signals , summing over , and dividing by for proper normalization. That is, we are led to conjecture that
This is in fact true, as we will later see. (It is straightforward to show as an exercise.)
FBS Window Constraints for R=1
STFT Filter Bank