---------------------------- Original Message ---------------------------- Subject: Re: [music-dsp] Auto-tune sounds like vocoder From: "Eder Souza" <ederwan...@gmail.com> Date: Thu, January 17, 2019 6:46 am To: "A discussion list for music-related DSP" <music-dsp@music.columbia.edu> -------------------------------------------------------------------------- > When I read the original patent US5973252A (Pitch detection and intonation > correction apparatus and method), in the vague description of how the pitch > shift is made, I wondered if everything seemed to be as simple as I would > be imagining... > > For pitch shift (auto Tune): > Just get the fracional period, use the fracional position period to cut off > or add periods whithout apply overlap and add (just splice and add/remove > in the exact period position, yeah this is why do you need a very strong > pitch detector, to join or discard in the exact period position to not get > clicks), this will expand or compress the signal, then now just resample > the signal to pitch shift (Ok now the formants go down). this is the standard kinda time-domain pitch shifting that goes by a variety of names: TDHS, maybe WSOLA.� it's what Eventide originally did (and i think what Autotune originally did).� when you splice out a period, that is time-compression (which speeds things up) and then for a pitch shifter, you have to resample that to slow down the time-compressed audio.� that moves both the pitch and the formants down.� for upshifting you are spicing in an extra period (which is time-stretching) and then resampling that to speed it up which moves the pitch and formants up.� in this method, the cycles of the quasi-periodic waveform or stretched or scrunched in the resampling.� and in this method octave errors might not hurt you because all that means is you splice in or out two entire periods, instead of one, and it's still a reasonably glitch-free splice.� this is particularly the case for WSOLA, which is not directly worried about the pitch at all, but in Waveform Similarity (which *is* indirectly related to pitch). this is different than the Lent/Hamon method (sometimes called PSOLA), in which you window off a single cycle and call it a "wavelet" or a "grain".� you do *not* stretch nor scrunch that wavelet or grain (unless you *do* wanna move the formants) but output the most current wavelet or grain, overlapping and adding, at the rate of the output pitch.� when upshifting, there is more overlapping and some grains will be used twice.� when downshifting, there is less overlapping and some grains will be skipped and not used. > > So I think that the current effect "Robotic or vocoded" happens when you > try change the formants (warping to original formants or just warping to a > new position). > > In the past I write the Keith Lent code to do Auto Tune(pitch correction) > and my results are cool... > > PS: I wrote to test the pitch detector described in the patent above (just > to proof of concept), and yeah this works great! well, i dunno where Autotune is now, but back in the '90s, i thought it sucked (the Wave Mechanics products, PurePitch and PitchDoctor were much better).� and Autotune back then did not use the Lent alg, so if there was a lot of shifting (like 3 or 4 semitones or more), Autotune definitely munchkinized the voice. -- r b-j� � � � � � � � � � � � �r...@audioimagination.com "Imagination is more important than knowledge." � � � �
_______________________________________________ dupswapdrop: music-dsp mailing list music-dsp@music.columbia.edu https://lists.columbia.edu/mailman/listinfo/music-dsp