[pulseaudio-discuss] New Module: module-lfe-lp

[Matthew Robbetts]

On 20 March 2013 03:48, Alexander E. Patrakov <patrakov at gmail.com> wrote:
> Matthew Robbetts <wingfeathera at gmail.com> wrote:
>> On 19 Mar 2013, at 18:34, "Alexander E. Patrakov" <patrakov at gmail.com>
>> wrote:
>>
>> > 2013/3/19 Justin Chudgar <justin at justinzane.com>:
>> > > I've created a module to ensure that only low frequencies are sent
>> > > to devices at the end of an "lfe"/"subwoofer" channel. This module
>> > > allows the user to select the master channel, the low pass cutoff
>> > > frequency (aka corner freq, -3dB freq) and the number of filter
>> > > poles.
>> >
>> > Sorry, I cannot sign off this filter implementation.
>>
>> The filter implementation looks fine to me. Do you mean the coefficient
>> calculation?
>
> I don't really know, that was a blackbox test. Will look again today.

Oh sorry, I thought you'd looked at the code.

>
>> > First, the filter currently fails the "attenuation must be 3 dB at the
>> > cut-off frequency" test that all Butterworth filters must pass.
>>
>> [That's not a very clear way to describe things. A filter is either a
>> Butterworth or it isn't, and that is determined by ts transfer function.
>> Specfic aspects of its frequency response are really side-effects of its
>> Butterworthiness.]
>
> Well, what is implemented is certainly not a Butterworth filter of Nth order with a given cut-off frequency.
>
>> Anyway: in what way does the filter not roll off correctly?
>
> I have not tested the roll-off. Only the statement about the -3 dB frequency.
>
>> The coefficient calculation looks to me like a quick swipe from the Audio EQ
>> cookbook (I haven't verified that). Those formulae work pretty well, so
>> if this filter isn't working right then presumably Justin has made a
>> small error in that somewhere.
>
> Quite possible. Or it may be that I have misunderstood the expected result.
>
> Looking again at the code, in do_filter() I see a strange loop over the poles that, essentially, filters the signal through the same biquad (and not through N/2 biquads with independent history) over and over again.

*sigh*, obviously. I missed that. The looping looks cute but yes, it
is doing it wrong. All the filters might share coefficients, but each
filter definitely needs its own biquad_data struct.



>
>> Justin, a relatively easy way to verify your coefficient generation is
>> to use Matlab's (well, Octave's!) 'butterworth' function. This will get
>> you the coefficients you need directly.
>>
>>
>> > Second, Q=sqrt(2) is only valid for a 2nd order Butterworth filter.
>>
>> Well, to be clear, it *makes* it a Butterworth filter. The Q in the
>> transfer function is the only thing distinguishing a Butterworth from
>> any other filter you can make with a biquad.
>
> Higher-order Butterworth filters use more than one biquad, with different values of Q. But indeed, a 2nd order Butterworth filter takes one biquad with Q=2.
>
>> Although, since you want him to use a cascade of Butterowrth filters, he
>> surely *should* use Q=2?
>
> Yes, for Butterworth filters of 2nd order. However he wrote some code that pretends (if I understood the intent correctly) to do arbitrary order lowpass Butterworth filtering with a given cut-off frequency.

Ah sorry, I see what you mean. Yes, you're right. Either, the
coefficient design code should be changed to the arbitrary filtering
correctly (probably quite a lot more code to do this), or the the user
interface should be constrained to only actually allow the user to ask
for second-order Butterworth. I agree that the latter is probably more
suitable.



I clearly have not added very much here. Sorry for the noise!