Hi, I am comparing coefficient quantization errors for second order IIR filter structures (for fixed point implementation). I have found the Chamberlin state variable structure to perform well at low frequencies and have seen in the article "The modified Chamberlin and Z�lzer filter structures" that the quantized pole distribution is dense in the low frequency area of the unit circle in the z-plane. I have implemented the filter (in software) based on the diagram in: http://www.earlevel.com/Digital%20Audio/StateVar.html Now I want to derive the transfer function from the diagram and find out how f and q (see the link) are related to the b-coefficients in: H(z) a0 + a1*z^-1 + a2*z^-2 = ------------------------ 1 - b1*z^-1 - b2*z^-2 Then I will find the relation to the poles and make my own quantized pole distribution plot. From the diagram I can see that the following simple implementation can be used (working matlab code, only HP filteret samples are saved): low = 0; high = 0; band = 0; y = zeros(length(x),1); for i = 1:length(x) low = low + f * band; high = x(i) - low - q * band; band = f * high + band; y(i) = high; end I have then tried rewriting it to something like this: ylp(n) = ylp(n-1)+f*ybp(n-1) yhp(n) = x(n) - ylp(n)-q*ybp(n-1) ybp(n) = f*yhp(n)+ybp(n-1) but the problem is I now have three equations with more than three unknowns and also I need to get some x[n-1] etc. into the equation. I have tried doing some substitution and for instance turned up with: ybp(n) = -f*ylp(n-1)+f*x(n)-(f^2+f*q-1)*ybp(n-1) Then by looking at the diagram I tried to express ylp(n-1) by some by some x(n-...), f and q's but as it is recursive I can't (don't know how to) do that. There must be some mathematical trick or sequency knowledge I should know? Best regards, Kim
Chamberlin state variable filter transfer function
Started by ●March 29, 2009
Reply by ●March 29, 20092009-03-29
kimtherkelsen@gmail.com wrote:> I have implemented the filter (in software) based on the diagram > in: http://www.earlevel.com/Digital%20Audio/StateVar.html > > Now I want to derive the transfer function from the diagram and > find out how f and q (see the link) are related to the > b-coefficients in: > > H(z) a0 + a1*z^-1 + a2*z^-2 > = ------------------------ > 1 - b1*z^-1 - b2*z^-2> ylp(n) = ylp(n-1)+f*ybp(n-1) > yhp(n) = x(n) - ylp(n)-q*ybp(n-1) > ybp(n) = f*yhp(n)+ybp(n-1) > > but the problem is I now have three equations with more than > three unknowns and also I need to get some x[n-1] etc. into the > equation.Pass to the z-domain to get a single variable for each signal. Omitting arguments in L = L(z) etc. for brevity, your equation system becomes L = L/z + f*B/z H = X - L - q*B/z B = f*H + B/z. Note that in this form, the highpass transfer function is not H but H/X. Isolate L and B and eliminate them from the second equation: L = f*B/z/A, A = 1-1/z = (z-1)/z B = f*H/A; H = X - B/z*(f/A + q) = X - H*f/A/z*(f/A + q). From that, isolate H, expand the polynomials, and finally augment the fraction by 1/z^2 to identify the coefficients: H = X / (1 + f/A/z*(f/A + q)) = X * (z-1)^2/(z^2 + (f^2 + f*q - 2)*z + 1-f*q), so a = (1,-2,1) and b = (1, 2-f^2-f*q, f*q-1). Martin -- Sufficiently encapsulated magic is indistinguishable from technology.
Reply by ●March 29, 20092009-03-29