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

Re: AUDITORY Digest - 25 Apr 2008 to 26 Apr 2008 (#2008-92)

Dear All,

The situation is much worse that Dick has indicated.
The cochlea is nonlinear, and the filter bandwidths change, not only on level but the type of stimuli, and specifically the onset of the signal.
Most of the psychophysical, BM and neural data are for tones. And the way the tones are gated on is not always controlled or specified. 

Now the signal we really care about is speech.
The onsets of speech seem to be controlled carefully, tuned to perception. I base this assumption on our speech perception research, which has shown that onsets are critically important (May 2008 JASA article).
Speech perception research using synthetic speech contains no assumptions/controls of onsets, and is therefore likely misleading.
Most studies deal with the NL cochlea in terms of the "cochlear amplifier" which is styled after a linear system that slowly changes it gain. Most do not even think of the cochlear amplifier as a NL system. I have never seen a model or study that worries about onsets.
Research on onsets on the BM or in the neuron, show some very strange effects. For example, on the BM the first half cycle is linear, and in the neuron, it is highly nonlinear.
Speech perception requires these onset dynamics, as discussed for the last 10 years in Heil and Neubauer's research.
In music, we know that onsets are critical. Yet speech research, light years behind, has yet to identify onsets as critical.
Please forward to the list (and CC me), any and all studies that have identified onsets as critical to speech perception. I would like to compile a complete list of such studies. 

Jont Allen

AUDITORY automatic digest system wrote:

There are 2 messages totalling 101 lines in this issue.

Topics of the day:

  1. ERB at high levels (2)

----------------------------------------------------------------------

Date:    Fri, 25 Apr 2008 20:13:42 -0700
From:    Pierre Divenyi <pdivenyi@xxxxxxxxx>
Subject: Re: ERB at high levels

Sorry, folks. I was sure that I got a glimpse of Dick Lyon's model from the
web -- apparently I was wrong. Just trust me: it does exist.
Dick: my sincere apologies!

Pierre

------------------------------

Date:    Fri, 25 Apr 2008 22:44:39 -0700
From:    "Richard F. Lyon" <DickLyon@xxxxxxx>
Subject: Re: ERB at high levels
Thanks, Pierre, but I don't think I have a distribution of the latest model and the latest filter fitting data yet, and it might not be the answer he's looking for anyway. 

The question as posed leads to two immediate difficulties:
1. Nobody knows exactly what the filters, or their shapes, are; and the ERB depends on what shape you assume, and what data you fit it to.
2. The dependence on dB SPL is ambiguous. What level should one use to parameterize the filter shapes? The level of a probe tone that's just audible in a simultaneous masker? The masker level per Hz? Other?
With help from Unoki and Patterson, I recently redid some fits to simultaneous notch-noise masker data of Baker et al. and of Glasberg and Moore, using several gammachirp, gammatone, and related filter models, including the pole-zero filter cascade (PZFC), which is probably what Pierre is referring to. I parameterized the ERB in terms of the threshold probe level, up to 70 dB SPL, in which case the total noise level is probably around 90 dB SPL, at least for some notch widths. At these levels the ERB is typically about twice the reference ERB curve of Glasberg and Moore 1990.
This is not significantly different from what Patterson, Unoki, and Irino reported in their 2003 JASA article. But the dependence of ERB on level does vary quite a bit depending on which filter model you choose to fit, so there's really not a definitive answer from these data sets and fits.
Whether the ERB becomes constant at high levels, or continues to grow, is not yet clear. And even if you get a clear answer for that from other psychoacoustic experiments, it doesn't necessarily mean that's exactly what's happening in the cochlear mechanics; or vice versa. But probably others do have more data points relevant to the question.
As for my latest PZFC cochlea model, it's now integrated into Roy Patterson's AIM-MAT package, using parameters determined by the fitting to the human masking data (under the assumption that the psychoacoustic features correspond to cochlea channel filters, which Steve Greenberg warns me may not be a good assumption). Hopefully an AIM-MAT release with this is not too far off; we may want to tweak it a bit more first. 

Dick


At 6:06 PM -0700 4/25/08, Pierre Divenyi wrote:

Ken,
Check out Dick Lyon's new cochleogram (also misnamed cochleagram) model on his web site. He spent a long time trying to come up with the correct parameters. In any case, the model is beautiful. 

Pierre

At 10:29 AM 4/25/2008, Ken Grant wrote:
We are looking for any data or model of ERB as a function of level. We are aware of the Oxenham and Simonson (2006) and the Baker and Rosen (2006) papers which don't go above 70 dB SPL. We would like to hear any suggestions as to what people think normal-hearing ERBs at 500, 1000, 2000, and 4000 look like at input levels as high as 90-95 dB SPL. 

--
Ken W. Grant

Walter Reed Army Medical Center
Army Audiology and Speech Center
Building 2, Room 6A53C
Washington, DC 20307-5001

Work: 202-782-8596
Fax: 202-782-9228

email: grant@xxxxxxxxxxxxx


------------------------------

End of AUDITORY Digest - 25 Apr 2008 to 26 Apr 2008 (#2008-92)
**************************************************************