rel4.2.doc

来自「speech signal process tools」· DOC 代码 · 共 661 行 · 第 1/2 页

ESPS Version 4.2 Release Notes                        page 1


               ESPS Version 4.2 Release Notes

                Document version: 1.4 2/1/93

1.  INTRODUCTION

This document provides notes for release 4.2 of ESPS.  These
release  notes  document  changes to ESPS since the revision
4.1.  This document can be viewed conveniently  using  einfo
(1-ESPS) or eversion (1-ESPS).

Note that when shipped as an update (i.e. not to new  custo-
mers)  new  hardcopy  manuals are not provided.   The online
documentation is up to date.  Please  refer  to  the  online
manual  pages  for  all  programs and library functions men-
tioned below.

2.  SUMMARY OF MAJOR CHANGES

This release is primarily a maintenance release and most  of
the changes were bug fixes to programs and error corrections
to the documentation.  Major highlights include:

   + major improvements to the filter design  and  filtering
     programs

   + license manager startup delay reduced from 3 minutes to
     1 minute

   + automatic checkout of ESPS license when required

   + e2sphere a license-free program added to  convert  ESPS
     files to Sphere header files

   + updated documentation

3.  CHANGES TO USER-LEVEL ESPS PROGRAMS

3.1.  CHANGES TO FILTER PROGRAMS

In the ESPS 4.2 release, the  numerical  stability  of  many
filter programs is improved.  The new filter2 program imple-
ments an IIR filter in cascade form  of  second  order  sec-
tions,  rather  than in direct form implementation as in the
old filter program. The filt_spec  and  impulse_rs  programs
have  been re-written for numerical stability for high order
IIR filters.

The    library    functions    block_filter2(3-ESPS)     and
init_fdata(3-ESPS),  the  core functions of filter2, replace
the old block_filter(3-ESPS) in the library.   The  function
fil_spectrum(3-ESPS)  is  added to the library for computing
filter spectral characteristics.










ESPS Version 4.2 Release Notes                        page 2


Chebyshev2 and elliptical  IIR  filter  design  methods  are
added  in  the  iir_filt  program.   The  user-interface  of
iir_filt is made easier than the previous release by requir-
ing  only the input parameters of passband and stopband fre-
quencies and loss characteristics.

A binary sfconvert program has been written to  replace  the
old  sfconvert  shell  script for sampling frequency conver-
sion.  The new program allows user to have complete  control
over  trade-off  between  computation efficiency vesus accu-
racy.

wmse_filt now stores correct number of bandedges in the ban-
dedges generic header item.

3.1.1.  SPECIFIC CHANGES TO FILTER PROGRAMS

(1)  iir_filt - Major IIR filter improvements:

     Elliptical and  Chebyshev II are now supported.

     Only need  to  specify  stop/passband  frequencies  and
     stop/passband loss for filter design.  No need to esti-
     mate the order, no need to know what 3dB frequency (for
     Butterworth),  or  critical frequency (for Cheveyshev1)
     is, if the user is not  familiar  with  filter  design.
     The  new  iir_filt  allows  a precise control on filter
     characteristics.

     The program computes the optimal filter order.  But the
     user  has  an option of over-riding the filter order by
     means of supplying his own filter order in the  parame-
     ter file.

     No array parameter  input.   This  means  you  can  run
     iir_filt  with xwaves by the use of xeparam to design a
     filter and process the data on the fly.

     Maximum passband gain is normalized to  0dB.   The  old
     iir_filt  has  its maximum passband amplitude different
     from 0dB if the filter order is odd.

(2)  wmse_filt - this  program  now  checks  for  a  license
     checkout  before  the user enters the parameters.   The
     header item nbands has been fixed to store the  correct
     number of bandedges.

(3)  filter2 - a new  numerically-stable  filtering  program
     intended to replace the existing program filter.

(4)  sfconvert - this is an all new program (the old  script
     by  the  same  name  is gone).   It has a number of new
     options and is much  faster  than  the  previous  shell










ESPS Version 4.2 Release Notes                        page 3


     script  version.    See the man page for details.  This
     program is also linked to the name esfconvert to  avoid
     a name clash on SGI machines by an SGI supplied program
     named sfconvert.

(5)  xsfconvert - XIG version of the new sfconvert.

(6)  impulse_rs - improved the numerical stability.

(7)  filtspec - improved the numerical stability.

(8)  plotspec - improved the numerical stability.

3.2.  CHANGES TO OTHER PROGRAMS

(1)  vqasc - this program has a new option -l to specify the
     maximum  length  of  input lines (removing a hard-coded
     limit).

(2)  vq - no longer limits the maximum codeword dimension to
     1000.  Allocations are done dynamically as needed.

(3)  lpcana - now uses anal_method correctly from the param-
     eter file.

(4)  exprompt - the output parameter file is now deleted  if
     exit was not the result of clicking on DONE.

(5)  me_spec - the parameters startrec and nrecs  have  been
     changed  to  start  and nan to be consistent with other
     ESPS programs.

(6)  fea_stats - the standard  output  has  been  made  more
     readable for non-stat files (changed output format from
     %f to %e). An error message has been added to  indicate
     if the covariance matrix is singular.

(7)  refcof - the -S0 option has been fixed.

(8)  psps - now prints the correct  bandedges  header  items
     for  FIR  FT_FILT files.  This program is linked to the
     name epsps to avoid a name clash with  a  Sun  supplied
     psps (part of Open Windows).

(9)  testsd - fixed sine drift problem.

(10) acf - this program now  correctly  computes  the  start
     time of the output file in cases where input files have
     negative start times.

(11) fftcep - this program now correctly outputs  both  real
     and imaginary parts of the complex cepstrum.











ESPS Version 4.2 Release Notes                        page 4


(12) fea_stats - this program computes an inverse covariance
     matrix     using    matrix_inv(3-ESPS).      The    old
     matrix_inv(3-ESPS) had a broken singular matrix  check,
     so  a  bad  inverse  covariance  matrix  was  generated
     silently in the fea_stats output.

(13) The following programs now place temporary files in the
     directory   specified   by   the  environment  variable
     ESPS_TEMP_PATH: acf, addfea,  auto  cross_cor,  dither,
     frame,   lpcana,   mergefea,  refcof,  select,  setmax,
     addgen, copysps, rem_dc, fea_stats, fea_edit, addclass,
     and comment.

(14) gpstohp - this is the GPS to HP Laserjet filter.   This
     has  been  improved  so that it produces smaller output
     files for a given plot.  In many cases the data sent to
     the  printer  has  been  cut  in half.   This filter is
     called by the plot programs when  the  -Thp  option  is
     used.

(15) plotsd - this program used to subsample (plot by simply
     skipping  points)  if  the  number  of  points  to plot
     exceeds the available number of pixels.  Now  the  pro-
     gram plots the min and the max of a segment.  This pro-
     duces plots that look just like xwaves.

(16) plotsd - fixed to  work  with  ESPS  common  processing
     turned off and removed the (now useless) -P option.

(17) playtest - fixed man page to mention that script  lines
     are  run twice.  Also the (nonfunctional) -n option has
     been removed.

(18) plot3d - after plot3d started up (on Suns), the initial
     graph  was offset and scaled differently than the space
     the cursor was moving in.  This was fixed by having the
     plot  frame size established BEFORE the enclosed canvas
     was created.

(19) addfea - this program failed on SGI machines  when  the
     -f option was not used.   It has been fixed.

(20) tofspec - The -s option was added to  allow  specifica-
     tion  of the generic header item sf in the output file.
     This is a generally useful change as it permits control
     over  the  frequency scale when the output is displayed
     via xwaves.  This change (when coupled with an  xspecav
     script  change) fixes the fairly serious xwaves display
     bug that shows incorrect frequency scale  for  spectral
     averages.   Note that there's a corresponding parameter
     as well (sf).  If it is not present or has  zero  value
     (also  the  default  of  -s),  then the old behavior is
     used.










ESPS Version 4.2 Release Notes                        page 5


(21) send_xwaves - Implemented and  documented  blocking  in
     send_xwaves.   This  permits  any  xwaves command to be
     issued reliably from send_xwaves without concern  about
     completion  of a previous command.  This new version of
     send_xwaves must be used with the new version of xwaves
     (3.1).   However, the old send_xwaves will work ok with
     the new xwaves.

(22) fft and fftinv -

     Now both -S0 on the command line and step =  0  in  the
     parameter  file  imply a step equal to the default; the
     frame length.

     Omitting the end of the range in the fft -r  option  is
     now  equal to setting nan = 0 in the parameter file and
     implies processing from the starting point through  the
     end of the file.

     The manual page has been fixed  to  document  nan  ade-
     quately.

     Previous version would not run on a pipe with nan =  0.
     The  main  problem  with  nan = 0 on a pipe is that the
     default frame length was the implied value of  nan:  in
     this  case  the  length  of the range from the starting
     sample to the end of the file.  That made it impossible
     to  allocate  storage for a frame of data without first
     reading the entire file.  We changed the default  frame
     length  to  be  equal to the transform length (2 to the
     power order), since that seemed like a more  reasonable
     default  in  any  case.  Now fft will run on pipes with
     nan = 0.

     fftinv start_time calculation has been fixed.

     fft scaling was wrong when frames overlap.  The scaling
     of  spectral  values  has been revised to be consistent
     with the total-power computation.  (The problem was not
     so  much with overlapping frames as with a frame length
     different from the transform length.)

     fftinv needed  compensating  change  when  fft  scaling
     fixed.   fftinv was changed to compensate for the scal-
     ing change in fft

     fftinv needlessly  requires  fft_length  generic.   Now
     fftinv  determines  the  fft  length from the num_freqs
     header item,  which  is  standard  in  FEA_SPEC  files,
     rather than from the fft_length item, which is not.  If
     the fft_length item is present, the program checks  for
     the  condition (fft_length == num_freqs - 1) and prints
     a warning in case of inconsistency.