; This is a collection of various "autotune" patches. "autotune" is an ; effect that pitch-shifts a monophonous input to the nearest halftone. ; ; Autotune is either useful to correct intonation problems (eg. for ; Karaoke or Instrumental tracks that might be out-of-tune by instrumental ; principle - a violinist playing solo eg. might get out-of-tune by a few ; cents during a longer piece due to lack of reference) or, with extreme ; settings ("strength" close to 1, decays low) as an effect that gives the ; sample a mechanical ("computer voice") feeling. ; ; The Song "Believe" by "Cher", which, despite being poison to your ears, ; was very popular in Germany, used an Autotune effect for distinctive ; chorus vocals. ; ; In the end, the effect is very simple: ; ; a) measure the input frequency f_i ; b) round it to the nearest halftone f_target ; c) pitch shift by the ratio (f_target/f_i) ; ; Refinements to this basic effect are: ; ; - ignore or dampen broken measurements ; - make f_target "glide" ; ; The devil is, like always, in the details. Neither measuring input ; frequency nor pitch shifting are particularly basic. ; ; There are two ways of measuring Frequency: by pvsanal ("V") and by pitchamdf ; ("P") and two ways of pitch shifting: using harmon ("H") or using a delay ; line ("D"). This gives four possible combinations of the autotune effect: ; ; AutotuneVD (which is the one i'd recommended) ; AutotuneVH ; AutotunePH ; AutotunePD ; ; Concerning the choice of methods: ; ; I found two satisfying ways to measure the Frequency of a sample. One uses ; pvsanal/pvspitch, which works fine even with very small FFT windows (which ; are necissary - the pitch correction has to work on a scale of under maybe ; 0.05 seconds - which rules out window sizes beyond 2048). However, it has ; the drawback that pitch can only be determined sensibly once a full FFT ; window has been received. Inbetween complete FFT windows, the measured pitch ; can not change. So a quick glide might not be corrected for. ; ; The other one uses pitchamdf, which uses a different method without such ; drawbacks. Additionally, it is computationally expensive. ; ; However, I think that both methods might be equally useful, so I ; included both. ; ; I also found two satisfying ways to pitch shift a sample. One is inspired ; by Hans Mikelsons pitch shifter (Instrument 22 in ; http://www.csounds.com/mikelson/multifx.orc), however improves on it by ; using two windowed "grains" and switching between them. The method is not ; free of artifacts, especially there's a periodic clicking noise. This might ; be because i don't fully understand the problem. ; ; The other one is simply using harmon, which works great, speaking in ; terms of reaction speed and signal delay, but sadly introduces some ; distortion. The artifacts are, however, different to the one produced by ; using delay lines. ; sr = 44100 ksmps = 1 nchnls = 1 opcode AutotuneVD,ak,akkk ifftsize = 512 iwtype = 1 asig,kstrength,kdecayin,kdecayout xin ; 0 <= kstrength <= 1 - 0: no effect 1: pitch shift to closest halftone as closely as possible ; kdecayin>0 (suggested: 0-3) error tolerance - higher values give additional decay time to target frequency, lower values make "off" frequency measurements break ; kdecayout>0 (suggested: 0-8) time-dependent decay of target frequency fsig pvsanal asig, ifftsize, ifftsize/4, ifftsize, iwtype kfr, kamp pvspitch fsig, 0.5 ; This is the one of the magical places: round to the closest halftone. ktarget = exp(0.021736895+round(12*((log(kfr))/log(2)))*log(2)/12.0) ; correct for errors, decay measured frequency: large decay times give better error tolerance but slower reaction ; measured frequency < 1 : invalidate (no motion at all) - 1 is pretty arbitrary ; Target frequency is weighted average of measured or true frequency ktarget = kfr*(1-kstrength)+ktarget*kstrength kfratiolp init 1 kfratiolp = (kfr<1)?kfratiolp:(kfratiolp*(1-exp(-kdecayin))+(ktarget/kfr)*exp(-kdecayin)) ; decay target frequency: decay that depends only on time kfratiodecay init 1 kfratiodecay = kfratiodecay*(1-exp(-kdecayout))+kfratiolp*exp(-kdecayout) ; use this for a really nasty fudge factor - overdo pitch shift - don't have a clue why this should be necissary, but it simply sounds better ;kpshift = log(kfratiodecay)*1.7 kpshift = log(kfratiodecay) ischlapp = 0.01 ; a fudge factor - initial delay if1 = 0.08 ; another fudge factor - delay line length kosc1 phasor kpshift,0 ; generate half sine at init time ihalfsine ftgen 0, 0, 128, 19, 0.5, 1, 0, 0 ; squared sine oscillator ksinsq tablei kosc1/if1,ihalfsine,1,0 ,1 ksinsq = ksinsq*ksinsq ; read two signals from the delay line, with half a delay line difference atmp delayr if1 a1 deltapi 1-kosc1 a2 deltapi 0.5-(kosc1>0.5?kosc1-1:kosc1) delayw asig ; blend out signals by aout = a1*ksinsq+a2*(1-ksinsq) xout aout,kfr endop opcode AutotunePD,ak,akkk asig,kstrength,kdecayin,kdecayout xin kfr, kamp pitchamdf asig, 40,600 ktarget = exp(0.021736895+round(12*((log(kfr))/log(2)))*log(2)/12.0) ktarget = kfr*(1-kstrength)+ktarget*kstrength kfratiolp init 1 kfratiolp = (kfr<1)?kfratiolp:(kfratiolp*(1-exp(-kdecayin))+(ktarget/kfr)*exp(-kdecayin)) kfratiodecay init 1 kfratiodecay = kfratiodecay*(1-exp(-kdecayout))+kfratiolp*exp(-kdecayout) kpshift = log(kfratiodecay) ischlapp = 0.01 ; a fudge factor - initial delay if1 = 0.08 ; another fudge factor - delay line length kosc1 phasor kpshift,0.25 ihalfsine ftgen 0, 0, 128, 19, 0.5, 1, 0, 0 ksinsq tablei kosc1/if1,ihalfsine,1,0 ,1 ksinsq = ksinsq*ksinsq atmp delayr if1 a1 deltapi 1-kosc1 a2 deltapi 0.5-(kosc1>0.5?kosc1-1:kosc1) delayw asig ; blend out signals by aout = a1*ksinsq+a2*(1-ksinsq) xout aout,kfr endop opcode AutotuneVH,ak,akkk ifftsize = 512 iwtype = 1 asig,kstrength,kdecayin,kdecayout xin fsig pvsanal asig, ifftsize, ifftsize/4, ifftsize, iwtype kfr, kamp pvspitch fsig, 0.5 ktarget = exp(0.021736895+round(12*((log(kfr))/log(2)))*log(2)/12.0) ktarget = kfr*(1-kstrength)+ktarget*kstrength kfrlp init 1 kfrlp = (kfr<1)?kfrlp:(kfrlp*(1-exp(-kdecayin))+kfr*exp(-kdecayin)) ktargetdecay init 1 ktargetdecay = ktargetdecay*(1-exp(-kdecayout))+kfrlp*exp(-kdecayout) ischlapp = 0.01 ; a fudge factor - initial delay if1 = 0.08 ; another fudge factor - delay line length aout harmon asig,(kfrlp<1?500:kfr),0.4,ktargetdecay,0,1,200,0.1 xout aout,kfr endop opcode AutotunePH,ak,akkk asig,kstrength,kdecayin,kdecayout xin kfr, kamp pitchamdf asig, 40,600 ktarget = exp(0.021736895+round(12*((log(kfr))/log(2)))*log(2)/12.0) ktarget = kfr*(1-kstrength)+ktarget*kstrength kfrlp init 1 kfrlp = (kfr<1)?kfrlp:(kfrlp*(1-exp(-kdecayin))+kfr*exp(-kdecayin)) ktargetdecay init 1 ktargetdecay = ktargetdecay*(1-exp(-kdecayout))+kfrlp*exp(-kdecayout) ischlapp = 0.01 ; a fudge factor - initial delay if1 = 0.08 ; another fudge factor - delay line length aout harmon asig,(kfrlp<1?500:kfr),0.4,ktargetdecay,0,1,200,0.1 xout aout,kfr endop instr 1 kf init 440 kf = kf*1.00001 ain oscili p4,kf,1 aout,kfr AutotuneVD ain,1,1,2 out aout endin 1 instr 2 kf init 440 kf = kf*1.00001 aout oscili p4,kf,1 out aout endin 2 instr 3 ain soundin "wolken-falsch.wav" aout,kfr AutotuneVD ain,1,1,2 out aout endin 3 f 1 0 1024 10 1 ; sine wave for sweep i 2 0 3 30000 ; play sweep i 1 4 3 30000 ; play autotuned sweep i 1 8 3 15000 i 2 8 3 15000 ; play sweep and autotuned sweep together ; listen for the beats! i 3 12 9 ; play "ueber den wolken" cher-style e