4
$\begingroup$

When using a phase vocoder to alter the spectral envelope of a signal, the input signal is split up into many (often hundreds of) bands. With many phase vocoders, the output levels these individual bands are often controllable by the user.

  1. What artifacts are caused by filtering these individual bands in a phase vocoder?

  2. If and why, phase vocoders are better suited (and cause less artifacts) than FIR filters (or other methods) for surgical changes to such narrow frequency bands. For example, are phase vocoders going to cause the same, or less, pass band ´ringing´ than using an FIR filter for the same task?

I understand phase vocoders only up to the point that they use an inverse DFT to resynthesise the sound, though Im not sure exactly how the individual bands are resynthesized, so I dont know what the problems/artifacts will be.

$\endgroup$
8
+50
$\begingroup$

The term "phase vocoder" may encompass more than one resynthesis method. One method uses a pruned frequency reassignment of the DFT basis vectors, and thus resynthesizes sound using a sparser set of basis vectors with slightly different frequencies than were present in the original DFT analysis windows.

This sparse basis may not be able to recreate a full spectrum wide-band signal, such as produced by a transient, as well as the original non-reassigned full DFT basis.

Phase vocoder analysis also makes the assumption that the spectrum is fairly stationary between FFT frames, an assumption that can also be violated by transients.

One term for the artifacts produced by this type of analysis/resynthesis is transient smearing.

$\endgroup$
  • $\begingroup$ Thanks. How about ´ringing´ artifacts, and spectral leakage specifically? I imagine that if an FIR were used for the same task, the quick rolloffs necessary for such narrow bands would mean ringing in the passbands, and any more gradual rolloffs would mean bands crossing over each other. Is ringing less of a problem in phase vocoders than FIR filters/other methods? $\endgroup$ – Dale Newton May 6 '12 at 0:18
  • $\begingroup$ Pruned spectral reassignment throws away information in some adjacent bins. This is great, as it eliminates most leakage. This is terrible if those adjacent bins, in addition, contain important parts of the signal. $\endgroup$ – hotpaw2 May 6 '12 at 0:27
  • $\begingroup$ Phase vocoder resynthesis is usually done using successively offset frames. This is great, as it eliminates almost all ringing past one or two frame times. This is terrible because it may cause an entire frame time to ring. Depends on where the transient start/end occurs with respect to the nearest frame boundary. $\endgroup$ – hotpaw2 May 6 '12 at 0:30
  • $\begingroup$ When information in adjacent bins is tossed, does this tend to result in the destruction of transients only, or is it often the case that tonal information, such as vibrato / harmonics etc are also lost (ie. Is this what you refer to when you say the effect of tossing information from adjacent bins can be ´terrible´?) $\endgroup$ – Dale Newton May 6 '12 at 2:15
  • $\begingroup$ regarding resythnthesis with offset frames - Im not sure what that means in terms of how it might sound. Taking frames of 60 milliseconds in length, does this mean that ringing would last possibly a whole frame (60 milliseconds), but no longer than one or two frames (120 milliseconds)? What would provoke this kind of ringing - a step response / other impulses, or just transients? $\endgroup$ – Dale Newton May 6 '12 at 2:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.