> The classic Phase Vocoder algorithm integrates the phase difference > between successive analysis frames to produce the phase component in > the synthesis frames. No explicit interpolation is needed.
To do time-stretching, you then have to multiply that unwrapped phase by the ratio of the hop sizes. Which is an explicit interpolation. E On Wed, Aug 20, 2014 at 2:13 AM, Joel Ross <joel.binarybr...@gmail.com> wrote: > The classic Phase Vocoder algorithm integrates the phase difference > between successive analysis frames to produce the phase component in > the synthesis frames. No explicit interpolation is needed. This is not > the same as results achieved by time domain methods. It seems that > reconstruction is never perfect in this case even for a one to one > mapping of input frames to output frames, though this could be an > error on my part - but I seem to remember that this is the case. I'd > be interested in an explanation as to why... > > Regards, > Joel > > On 20 August 2014 02:26, Ethan Duni <ethan.d...@gmail.com> wrote: > > It's been a while since I've thought about phase vocoders, and I've never > > gotten real hands-on in building a time stretch algorithm, so let me make > > sure I understand what you're doing: > > > > On the analysis side you've got a fixed frame size and a fixed hop size. > On > > the synthesizer side, you have the same frame size but a variable hop > size > > (depending on stretch factor). Right? Is there are processing of the FFT > > coefficients, or are they simply fed directly into the IFFT on the > > synthesizer side? If you are not doing any frequency-domain processing, > > then what you have is really a granular synthesizer. You can simply > > eliminate the FFT and IFFT, since they do nothing except introduce > > numerical error. This is simply a time-domain system that slices the > input > > into (overlapping, windowed) chunks and then re-spaces said chunks to > > achieve the effect. > > > > Unless there is also frequency domain processing going on? Then that's a > > different story. > > > > Granular synthesis based time modification is straightforward to > implement, > > low-complexity, and works okay for small modifications of the time scale. > > But at some point its granularity becomes pronounced - for a large enough > > stretch factor, your synthesis frames will no longer overlap, and so no > > amount of monkeying with the shapes of the windows is ever going to get > you > > to anything like a perfect reconstruction state. Which can be a cool > effect > > in its own right, but it *is* an effect. > > > > And that's okay, if what you want to achieve is a decent, cheap time > > stretch algorithm for small time-scale modifications. A bit of window > > manipulation or dynamic scaling will mostly address they modulation > effects > > you get in that range. Also the algorithm will be pretty cheap, since you > > can eliminate the FFT and IFFT. > > > > The way to do a serious phase vocoder based time stretch is to leave the > > synthesis hop size fixed (and equal to the analysis hop size). That way > the > > whole chain is known to satisfy the perfect reconstruction condition. > I.e., > > as long as you don't monkey with the FFT coefficients, you'll get out the > > same signal as you put in (modulo some numerical error). To do time > > stretching, then, you interpolate the frequency domain data. > Specifically, > > get the FFT coefficients in polar form, and then separately interpolate > > their phases and magnitude (probably in the dB domain for the latter). > This > > will work pretty well for tonal signals, but will fail for noise-like > > content. This is because the phase of noise-like content doesn't evolve > > smoothly, but rather jumps around randomly. So interpolating it will > impose > > an artificial coherence on the noise. A high-performance phase vocoder > time > > stretch algorithm will employ some kind of classification stage, to work > > out which FFT bins are coherent signals and which are noise-like, and > then > > apply appropriate interpolation to each of them separately. > > > > E > > > > > > On Tue, Aug 19, 2014 at 1:49 PM, Daniel Varela < > danielvarela...@gmail.com> > > wrote: > > > >> Thanks for your quick reply Ethan. > >> > >> Indeed I was missing some relevant information on my first email. I am > >> using the FFT / IFFT approach instead of the filter bank one. So I made > my > >> FFT frame size and analysis hop size fixed, achiving time stretching by > >> changing the synthesis hop size by the time stretch ratio. > >> > >> My implementation is something very similar to the one exposed here > >> <http://www.ece.uvic.ca/~peterd/48409/Bernardini.pdf>, but modified to > >> work > >> in real time processing frames of fixed size. > >> > >> > >> 2014-08-19 21:30 GMT+01:00 Ethan Duni <ethan.d...@gmail.com>: > >> > >> > Maybe I'm missing something obvious, but shouldn't the filter bank > itself > >> > be constant? I.e., no change in the overlap or windowing. The time > >> > stretch/compression is obtained by extrapolating/interpolating the > >> analysis > >> > parameters, not by shifting around the synthesis filter bank relative > to > >> > the analysis filter bank. > >> > > >> > No? > >> > > >> > E > >> > > >> > > >> > On Tue, Aug 19, 2014 at 12:42 PM, Daniel Varela < > >> danielvarela...@gmail.com > >> > > > >> > wrote: > >> > > >> > > Hello to the list! > >> > > > >> > > I am working in a time stretching effect using a standad phase > vocoder > >> > > approach. > >> > > > >> > > Basically I have an input parameter to control the time stretch > ratio, > >> > > which can change in real time from 0.5 ( sinthesys hop size = 1/2 > >> > analysys > >> > > hop size, x200% original time stretch) to 2.0 > >> > > ( sinthesys hop size = 2 analysis hop size, x50% original time), > going > >> > > through all intermediate values. > >> > > > >> > > The problem I am facing is that when the sum of the products of > >> analysis > >> > > and synthesis windows is not equals to unity, I get, as expected, > >> > > modulation on the final reconstructed signal. > >> > > > >> > > I have tried with the usual windows (Hanning, Hamming, Truncated > >> > Gaussian, > >> > > etc ) and each of then achieves perfect reconstruction for different > >> > > overlapping factors, but as soon as the time stretch ratio changes > the > >> > > modulation appears. > >> > > > >> > > The paper FAST IMPLEMENTATION FOR NON-LINEAR TIME-SCALING OF STEREO > >> > SIGNALS > >> > > < > >> > > > >> > > >> > https://files.nyu.edu/jb2843/public/Publications_files/Ravelli-DAFx-2005.pdf > >> > > > > >> > > discuss this problem (4.2), suggesting to use assymetric triangular > >> > windows > >> > > for the synthesis. > >> > > > >> > > Anyone else has faced this problem? > >> > > > >> > > Thanks in advance. > >> > > > >> > > Daniel > >> > > -- > >> > > dupswapdrop -- the music-dsp mailing list and website: > >> > > subscription info, FAQ, source code archive, list archive, book > >> reviews, > >> > > dsp links > >> > > http://music.columbia.edu/cmc/music-dsp > >> > > http://music.columbia.edu/mailman/listinfo/music-dsp > >> > > > >> > -- > >> > dupswapdrop -- the music-dsp mailing list and website: > >> > subscription info, FAQ, source code archive, list archive, book > reviews, > >> > dsp links > >> > http://music.columbia.edu/cmc/music-dsp > >> > http://music.columbia.edu/mailman/listinfo/music-dsp > >> > > >> -- > >> dupswapdrop -- the music-dsp mailing list and website: > >> subscription info, FAQ, source code archive, list archive, book reviews, > >> dsp links > >> http://music.columbia.edu/cmc/music-dsp > >> http://music.columbia.edu/mailman/listinfo/music-dsp > >> > > -- > > dupswapdrop -- the music-dsp mailing list and website: > > subscription info, FAQ, source code archive, list archive, book reviews, > dsp links > > http://music.columbia.edu/cmc/music-dsp > > http://music.columbia.edu/mailman/listinfo/music-dsp > -- > dupswapdrop -- the music-dsp mailing list and website: > subscription info, FAQ, source code archive, list archive, book reviews, > dsp links > http://music.columbia.edu/cmc/music-dsp > http://music.columbia.edu/mailman/listinfo/music-dsp > -- dupswapdrop -- the music-dsp mailing list and website: subscription info, FAQ, source code archive, list archive, book reviews, dsp links http://music.columbia.edu/cmc/music-dsp http://music.columbia.edu/mailman/listinfo/music-dsp
