2. Extract the residual (what's left after you inverse-filter out the formants)
3. Use THAT as the excitation source instead of the mathematical LF pulse
4. The formant filters would shape it, but the "texture" would be real voice
It's basically what CELP/LPC vocoders do, but in reverse - instead of coding speech, you're generating it with a hybrid source.
Tools like Praat, STRAIGHT, or various MATLAB/Python libraries can do this. The result sounds like a weird buzzy/creaky "brrrrr" - not like speech at all, because all the "vowel-ness" has been taken out.

Sean | Ginsenshi The blindwolf
in reply to Tamas G • • •