[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: formant filtering

Dear Eckard,
  There was a most interesting coincidence concerning your remarks on
removing the "annoying highly tonal noise" caused by a pulsed arc welding
process. Before seeing your note I had conducted an experiment on the
pulsed arc waveform example (S1.wav) that you sent me just a year ago.
(Ref. M197 and M202) This might be the same example you referred to in your
note below. What happened was that while using this waveform to test my
method for identifying time-domain features I found a way to remove its
tonal components.
  As you may remember, I have been using my waveform information vector
(WIV) method to select and identify impulsive sounds. The purpose of this
method, as I have described it in previous communications, (M34) is to
circumvent the well-known difficulties inherent in resolving spectral
features of wideband impulses. At its basic level the WIV method can
separate impulses from reverberations and background sound. Recently I have
improved the algorithm to display in real time the semantic relationships
of various classes of impulsive sounds. (For example, "boom," "pop,"
"click," etc.)
  My system obtains identifiable features of impulses based on waveform
zeros. Each impulse is segmented at its onset and offset. Sequences of
zeros that represent meaningful waveform features are transformed into
two-dimensional patterns that are converted instantaneously into templates.
These templates are progressively collected in a multidimensional array
that relates classes of waveshapes to semantic meanings. The array method
permits instantaneous recognition of repetitions of impulse patterns.
  For an evaluation of the system's operation in a really difficult
impulsive sound environment I decided to use the impulsive complexity of
your welding sounds. The S1.wav was especially interesting because it
contains the "frying" sound of the welding process that is mixed with the
tonal features of the arc's periodic repetition.
  It was while running tests to see what portions of the sound would be
recognized when using templates having particular feature patterns that I
discovered I could remove the tonal sounds from the welding sounds.
  This experiment, among many others, has shown that components of a
complex acoustic environment can be selected and separated in the time
domain. It also demonstrates that separating sounds by the WIV method
avoids the limitations of spectral bandwidth.
  If tonal annoyance is still a pervasive problem in arc welding this
method might be of value in removing such interference. Conversely, it
could be useful for automatically monitoring particular sounds of the
welding process.
  I regard these experiments on impulses as the starting point for
developing a way to recognize, classify, and separate sounds in stages of
increasing degrees of complexity which could culminate in truly robust
speech recognition. My plan has been to get useful results by _not_ trying
to imitate cochlear or neural functions. But that's another story. (See M115)

  John Bates

>This difficulty to have an adequate basis for comparison reminds me of own
>experience. We tried to use spectral smearing for mitigation of an annoying
>highly tonal noise which was caused from a pulsed arc due to precise power
>electronic switching. The method can avoid dangerous mechanical resonance.
>We were also able to make the acoustic noise less 'cutting'. However, this
>required to randomize frequency over a considerable range of deviation from
>its technologically required mean value. We achieved a satisfactory
>decrease of tonality only on condition this range exceeded the critical
>bandwidth, while in this case the noise did perhaps not get less annoying
>but, as expected, even louder.