Next |
Prev |
Top
|
JOS Index |
JOS Pubs |
JOS Home |
Search
We will now analyze the filter bank interpretation of the STFT with
hop size set to
.
The kth subband signal in the DFT filter bank can be written as:
with
.
The signal
is regarded as a complex sequence formed from the
th DFT bin over time
(in frames).
Making the change of variable
, the above equation becomes:
![$\displaystyle X_k(m) = \sum_{r=0}^{N-1}u_r(m)W_N^{kn} \protect$](img391.png) |
(11) |
with
![$\displaystyle u_r(m) \equiv \sum_{l=-\infty}^{\infty}h(mN-lN+r)x(lN-r) \protect$](img392.png) |
(12) |
Therefore, (
) can be interpreted as computing a
length-
FFT applied to the input block
.
Now, we'll form a simpler representation of the sequence
.
First, define the polyphase decomposition of
and
:
If the filter
has length
, then each of its polyphase
components
has length
.
Now define the sequence
Finally, (
) can be expressed as:
![$\displaystyle u_r(m) = \sum_{l=-\infty}^{\infty}p_r(m-l)x_r(m)$](img401.png) |
(13) |
Thus, every input bin to the DFT is actually a convolution between
and
, the
th polyphase filter of the lowpass
prototype filter,
.
- Complexity is now the cost of one full-rate convolution with
and the FFT cost.
- If
(
), we can keep aliasing error within tolerable limits.
- If
(
), it is possible to keep the reconstruction error
below 0.1%.
- Similar development for the synthesis DFT bank
- Notice that when the polyphase filters are scalars (1-tap FIR filters)
of unit gain, then this is simply a DFT block transform, with a
rectangular window. The frequency response is a sinc, with poor
frequency characteristics.
- By extending the length of the polyphase filters,
, then
the freqeuency characteristics of the window can become much better.
- The window length is no longer restricted to be of the same
length as the transform.
Next |
Prev |
Top
|
JOS Index |
JOS Pubs |
JOS Home |
Search
Download JFB.pdf
Download JFB_2up.pdf
Download JFB_4up.pdf
[Comment on this page via email]