Re: [Discuss-gnuradio] Question on IF filter design for FM demodulation

[Achilleas Anastasopoulos]

Eric Blossom wrote:
> On Thu, Jan 06, 2005 at 05:37:53PM -0500, Achilleas Anastasopoulos wrote:
>
> > Dear all,
> >
> > I am looking at the FM demodulator example
> > provided in the gnuradio examples (in the mc4020 directory)
> > and I have a question on the
> > the design parameters of the filter used to separate one
> > FM station before demodulating:
> >
> > the sampling frequency is 20 Msps
> > the desired station is centered at 5.75 MHz
> > the designed filter (lowpass version)
> > has a cutoff frequency at 250 KHz
> > and a transition band of 800KHz
> > The signal is filtered and IQ demodulated and then
> > decimated by a factor of 125.
> >
> > It seems that this filter is too wide:
> > an adjacent FM station (at 5.75 MHz + 200 KHz)
> > is not attenuated at all and its spectrum (after decimation)
> > is folded on the desired station spectrum.
> > This is clearly seen in the simple example I attach.
> >
> > What am I missing here?
>
>
> Nothing.  It is too wide, but it works (mostly).
>
> The windowed filter design routine generates a lot of taps, and it
> gets too expensive if we use the "right spec".  It's probably worth
> using the Parks McClellan code to build a better filter and see how
> many taps you end up with.
>
> If you look at the USRP examples (these may only be in CVS) they have
> the right width and sound much better.  They're doing the channel
> selection and downconversion in the FPGA, passing a 256kS/s complex
> signal across the USB for each FM station.
>
> Take a look at gnuradio-examples/python/usrp1/{wfm_rcv.py,wfm_rcv_many.py}
>
> Eric

I am amazed that even this design (wfm_rcv.py) works:

Before the downconversion at the FPGA the entire FM band (88-108 MHz) is 
present inside the useful spectrum (with the desired station centered at 
5.75 MHz).
Then the digital downconversion brings the desired station at 0.
So far so good.

However, decimating at the FPGA by a factor of 250 will essentially 
spread the entire spectrum of the desired station (not a problem),
but also all other (neighboring) stations will be aliased on top of the 
desired one. These stations will appear in the desired band completely 
unattenuated.

It seems to me that before decimation at the FPGA, a LPF is required
(probably in two stages as Matt suggested).
Can this be done in the FPGA of the USRP?


Achilleas