Formant Filtering Example

In speech synthesis [27,39], digital filters are often used to simulate formant filtering by the vocal tract. It is well known [23] that the different vowel sounds of speech can be simulated by passing a ``buzz source'' through only two or three formant filters. As a result, speech is fully intelligible through the telephone bandwidth (nominally only 200-3200 Hz).

A formant is a resonance in the voice spectrum. A single formant may thus be modeled using one biquad (second-order filter section). For example, in the vowel $[a]$ as in ``father,'' the first three formant center-frequencies have been measured near 700, 1220, and 2600 Hz, with half-power bandwidths^10.7 130, 70, and 160 Hz [40].

In principle, the formant filter sections are in series, as can be found by deriving the transfer function of an acoustic tube [48]. As a consequence, the vocal-tract transfer function is an all-pole filter (provided that the nasal tract is closed off or negligible). As a result, there is no need to specify gains for the formant resonators--only center-frequency and bandwidth are necessary to specify each formant, leaving only an overall scale factor unspecified in a cascade (series) formant filter bank.

Numerically, however, it makes more sense to implement disjoint resonances in parallel rather than in series.^10.8 This is because when one formant filter is resonating, the others will be attenuating, so that to achieve a particular peak-gain at resonance, the resonating filter must overcome all combined attenuations as well as applying its own gain. In fixed-point arithmetic, this can result in large quantization-noise gains, especially for the last resonator in the chain. As a result of these considerations, our example will implement the formant sections in parallel. This means we must find the appropriate biquad numerators so that when added together, the overall transfer-function numerator is a constant. This will be accomplished using the partial fraction expansion (§6.8).^10.9

The matlab below illustrates the construction of a parallel formant filter bank for simulating the vowel $[a]$ . For completeness, it is used to filter a bandlimited impulse train, in order to synthesize the vowel sound.

F =  [700, 1220, 2600]; % Formant frequencies (Hz)
BW = [130,  70,  160];  % Formant bandwidths (Hz)
fs = 8192;              % Sampling rate (Hz)

nsecs = length(F);
R = exp(-pi*BW/fs);     % Pole radii
theta = 2*pi*F/fs;      % Pole angles
poles = R .* exp(j*theta); % Complex poles 
B = 1;  A = real(poly([poles,conj(poles)]));
% freqz(B,A); % View frequency response:

% Convert to parallel complex one-poles (PFE):
[r,p,f] = residuez(B,A);
As = zeros(nsecs,3);
Bs = zeros(nsecs,3);
% complex-conjugate pairs are adjacent in r and p:
for i=1:2:2*nsecs
    k = 1+(i-1)/2;
    Bs(k,:) = [r(i)+r(i+1),  -(r(i)*p(i+1)+r(i+1)*p(i)), 0];
    As(k,:) = [1, -(p(i)+p(i+1)), p(i)*p(i+1)];
end
sos = [Bs,As]; % standard second-order-section form
iperr = norm(imag(sos))/norm(sos); % make sure sos is ~real
disp(sprintf('||imag(sos)||/||sos|| = %g',iperr)); % 1.6e-16
sos = real(sos) % and make it exactly real

% Reconstruct original numerator and denominator as a check:
[Bh,Ah] = psos2tf(sos); % parallel sos to transfer function
% psos2tf appears in the matlab-utilities appendix
disp(sprintf('||A-Ah|| = %g',norm(A-Ah))); % 5.77423e-15
% Bh has trailing epsilons, so we'll zero-pad B:
disp(sprintf('||B-Bh|| = %g',...
             norm([B,zeros(1,length(Bh)-length(B))] - Bh)));
% 1.25116e-15

% Plot overlay and sum of all three 
% resonator amplitude responses:
nfft=512;
H = zeros(nsecs+1,nfft);
for i=1:nsecs
  [Hiw,w] = freqz(Bs(i,:),As(i,:));
  H(1+i,:) = Hiw(:).';
end
H(1,:) = sum(H(2:nsecs+1,:));
ttl = 'Amplitude Response'; 
xlab = 'Frequency (Hz)';
ylab = 'Magnitude (dB)';
sym = ''; 
lgnd = {'sum','sec 1','sec 2', 'sec 3'};
np=nfft/2; % Only plot for positive frequencies
wp = w(1:np); Hp=H(:,1:np);
figure(1); clf;
myplot(wp,20*log10(abs(Hp)),sym,ttl,xlab,ylab,1,lgnd);
disp('PAUSING'); pause;
saveplot('../eps/lpcexovl.eps');

% Now synthesize the vowel [a]:
nsamps = 256;
f0 = 200; % Pitch in Hz
w0T = 2*pi*f0/fs; % radians per sample

nharm = floor((fs/2)/f0); % number of harmonics
sig = zeros(1,nsamps);
n = 0:(nsamps-1);
% Synthesize bandlimited impulse train
for i=1:nharm,
    sig = sig + cos(i*w0T*n);
end;
sig = sig/max(sig);
speech = filter(1,A,sig);
soundsc([sig,speech]); % hear buzz, then 'ah'

Notes:

• The sampling rate was chosen to be $f_s = 8192$ Hz because that is the default Matlab sampling rate, and because that is a typical value used for ``telephone quality'' speech synthesis.
• The psos2tf utility is listed in §J.7.
• The overlay of the amplitude responses are shown in Fig.9.6.

Figure 9.6: Overlay of section amplitude responses and their sum.