Bob Masta <N0Spam@daqarta.com> wrote:
>>Bob Masta <N0Spam@daqarta.com> wrote:
(snip)
>>> There are problems with envelope detection, and especially
>>> hair cell rectification, as explanations for the missing
>>> fundamental. Consider that the cochlea behaves like a bank
>>> of mechanical filters, essentially one for each frequency we
>>> can resolve. The (inner) hair cell is the detector *after*
>>> a filter, so electrical rectification there does not
>>> introduce any mechanical distortion product that can
>>> propagate to the filter for the missing fundamental
>>> frequency.
>>OK, but say you add up the signals from all the hair cells.
>>I don't know if it helps, but as I understand it, told to me
>>by someone who does actual experiments, that hair cells send
>>out nerve impulses at the peak of the sine for low frequencies,
>>and approximately every 2nd or 3rd, or maybe more, for higher
>>frequencies.
> True, considering the inner hair cell and the neuron it
> feeds as a system. (The IHC doesn't actually generate
> impulses, it has a "graded" response in releasing
> transmitter molecules onto a very long "dendrite" fiber,
> whose cell body is in the cochlear nucleus and is the neuron
> that actually generates the spikes.)
Yes. The actual reason I asked this years ago was the age old
question of phase. If the signals from the cochlea don't preserve
phase, then obviously we can't hear phase, but they do.
> If the system is near the firing threshold, then a peak in
> the incoming frequency is more likely to push it over into
> firing. Once it fires, a neuron goes into a refractory
> period while it recharges (pumps ions across its cell wall
> to get back well below threshold) before it can fire again.
> That's why its max firing rate is only a few hundred Hz.
> But then when it's again able to fire, it will be more
> likely on whatever peak arrives next. If you plot a
> histogram, you see spikes clumped at periods that match the
> tuned frequency, even though many sine peaks may be missed
> between sequential pulses.
>>Seems to me that with the right weighting and time averaging,
>>one might approximate the envelope.
> Yeah, but then after that weighting and averaging, you'd
> only be up to the point of having a putative *input* to a
> correlator that might hope to extract the periodicity.
Oh, yes, I was just saying that there could be an envelope
detection, not what you would do with one.
> Remember, the brain is looking for one-fiber-per-frequency
> inputs. To detect the "missing" fundamental, it needs a
> signal on the line that corresponds to that particular
> frequency, the same line that would be excited by the
> fundamental alone. So your correlator needs a separate
> output line for each detectable frequency.
Well, as I understand it the Q is pretty low, at least by DSP
standards, so many nearby frequencies will also have a signal.
At some point, it has to be processed such that we know it is
a single tone.
> But there is a *much* simpler way, which as far as I can
> tell nobody is considering (including hearing researchers).
> And it doesn't use any tricky circuits like correlators that
> nobody knows how to make with ordinary neurons. It's one of
> those head-smackingly obvious things, once you look at the
> system and its existing components.
> Consider a fundamental of (say) 110 Hz, which excites a
> neural pathway at a firing rate proportional to the
> amplitude of the fundamental. To activate that same
> pathway when the fundamental is missing and only harmonics
> are present, we need to detect the presence of activity on
> the 220 Hz line AND the 330 Hz line, and various
> combinations of higher harmonic lines. Those lines are
> right there, we just need a few "analog AND gates" to
> combine them.
At some point, there is a system that does pattern recognition.
Is this a sound that I have heard before? But the sounds won't
have exactly the same spectral content, so it has to be able
to match them even a little different. That probably works better
for more natural sounds, such as human voice, and less for sums
of a small number of harmonics, though. Seems that it wouldn't
be hard to test different combinations of harmonics at different
amplitudes to see though.
> A single neuron is perfect for this; that's what it normally
> does. It sums/integrates all input synapse contributions,
> and if-and-when the sum reaches threshold, it fires. To get
> an AND effect, the proportions of input synapses ony need to
> be such that the (near) simultaneous contributions from 2 or
> more harmonic lines are needed to reach threshold. (An OR
> is when each input line has enough moxie to reach threshold
> on its own.) The integration period (volume of ions that
> need to be pumped out of the cell to reach threshold)
> controls how "simultaneous" the harmonics need to be.
It is supposed to be that babies learn pretty young which
tones are important in the spoken languages that they hear.
Past some age, maybe about one, those patterns are pretty
much fixed. Since human voice has pretty many harmonics,
that would go into the tone detection.
> Now, you'd be right in thinking that this scheme will have
> trouble working at different signal levels: After all, if
> the spike rates are doubled, a 2-input AND would fire with
> only one input active. And if all spike rates are halved,
> it might not fire even with both inputs active. I think
> this can be compensated fairly easily using an inhibitory
> input to act as a sensitivity control.
But it has to work at different levels. We recognize people
by their voice at varying levels. Even worse, our spectral
sensitivity changes with amplitude, the reason for loudness
controls on audio systems.
> Yes, this scheme would take several extra neurons for each
> of the few thousand resolvable frequencies. To the brain,
> that's chump change. And I suspect it is far cheaper than
> any sort of correlation scheme.
I think everyone with a cat (and probably dogs, too) knows how
fast they learn the sound of a can opener. (We don't give our
cats canned food very often, and they still learn it.)
> This approach seems like "Duh!"-obvious, which might mean
> that there is something "Duh!"-stupid wrong with it. So I'm
> depending on you guys to swat it down to size!
-- glen