🚀 go-pugleaf

In article <952@eplunix.UUCP> raoul@eplunix.UUCP (Nico Garcia) writes:
>I also had a fascinating conversation with an audiologist here about
>different compression schemes. I won't go into details since I'm an EE, not
>an audiologist, but one scheme that works surprisingly well is clipping:
>map a 60 dB range speech signal into a 30 dB comfortable range for the
>subject by clipping the upper 30 dB right off.

Actually, one-bit speech is intelligible: take your digital speech,
map all positive numbers into the same value, and all negative numbers
into the same value (with opposite sign).  Sounds like hell but you can
understand it.

>I believe that more effective processing for speech enhancement will
>be based on what humans actually do, rather than what our mathematical
>analyses insist is most efficient or powerful. I'm not certain, however,
>that we understand enough of how people *hear* and process sound to 
>approach speech enhancement from the speech production side.

People who do real speech processing fight this battle from both ends,
and while there are many mysteries it's not as black an art as you
suggest that it is.  For example, in speech compression your ears really
don't care if your method has the minimum mean-square error; however, if
you apply a certain type of frequency weighting before you attempt to
minimize the error it works quite well.  To produce low-bit-rate speech
of good quality you have to know a lot of DSP and ALSO know a good deal
about the features of the human auditory system.


--
--
Joe Buck
jbuck@galileo.berkeley.edu	 {uunet,ucbvax}!galileo.berkeley.edu!jbuck