COST250 Speaker Recognition Reference System
History
Version 1.0, 20/12/1999
Changes from 1.0b4:
- Minor editorial changes to documentation.
- Removed
-markers in documentation.
Version 1.0b4, 3/12/1999
Changes from 1.0b3:
- Changing default (reference!) codebook size from 16 to 64. Note: this
will change the results from previous (beta) versions of the Reference
System.
- The Reference System now unpacks into a directory named like the
version of the Reference System. Documentation updated accordingly.
- Changed web address for latest
version of the Reference System.
- 'make' or 'make info' now lists available make
targets.
- Created scripts that compute the EER (eer.csh that uses eer.awk). They can be installed on the
system with make.
- 'make test' now computes the EER and creates the data
file for a DET-plot after the experiment has been run.
- 'make compare' now also compares the DET data file with
a reference file. Updated documentation accordingly. The
DET-plot and EER scripts are new and their portability has not yet
been tested. Please report problems to melin@speech.kth.se.
- Added links to source code files referred to in the text.
- Added a new section with recognition
results.
- Added a new section with a list of
files in the distribution and links to them.
- Added PowerPoint and HTML versions of a presentation of the
reference system.
Version 1.0b3, 8/11/1999
Changes from 1.0b2:
- All source code to the reference system is now distributed with
documentation. See the copyright notice
for terms of copying and usage.
- Numbering sections in the documentation.
- Preliminary scoring software has been added:
- Can still do only the test set ROC and not compute EER.
- Scripts are installed by makefile.
- Documented in section 7.
Version 1.0b2, 26/10/1999
Changes from 1.0b1:
- Documentation:
- Added references for WG4, EAGLES
handbook, VQ and world model score normalization.
- Marked those engine properties in Table 2 that must not be changed
for the system to perform as the reference system.
- Updated copyright regulations.
- Minor formatting changes.
- Other minor changes.
Version 1.0b1, 21/06/1999
Changes from 0.2:
- The default LPC configuration options have been changed from
"-f200 -o30 -p8 -s1" to "-f200 -o60 -p12 -a0.97 -s1". The new setting
chooses 12th order LPC with 100 frames per second computed from a 25
ms analysis window (200 point windows with 60% overlap), assuming the
input signal has 8000 Hz sampling rate. The preemphasis coefficient
is now 0.97 instead of 0.95.
- It is now possible to specify whether score normalization shall
be used or not, using the "normalization"
property in the ReferenceSystem
class. If normaliation is enabled, a set of non-client models can
also be specified using the "nonClientModels" property. For each
verification test, the model in the set with the smallest distance
(highest score) is chosen for score normalization. This feature can be
used for instance to specify a pair of gender-dependent non-client
models instead of a single world model. Default is to use a single
world model called "W".
- Experiment definition files for version 2.0 of Polycost baseline experiments
are now included in
experiment/polycost/be[1-3]/v2.0.
To run a Polycost baseline experiment, type 'make beX'.
- A new chapter in the documentation that describes the operation
of the reference system. Some minor changes to the documentation have
also been made.
-
string.h is now included where needed (avoids warnings when
compiling on IRIX with cc).
- A Windows style zip-file with the Polycost subset needed for the calibration
test is now available here.
- A bugfix in
src/concat.c: the main function
previously had a return type void which caused different
exit codes on different platforms. It now has an int
return type and returns 0 on successful operation.
Version 0.2, 24/02/1999
Changes from 0.1:
- Installation notes have been improved to better describe how to install
the Tcl package cost250.
- Added error detection and presentation. Previously most errors in the Tcl code were not
caught and resulted in an uninterpretable trace dump.
- The database object is now flexible with upper and lower case letters in
the tag and filename extension when looking for a file in the database. If the
file is not found with a name exactly as given by tag and extension, it tries
several combinations of upper and lower case conversions. This feature is useful
in connection with files stored (or copied from) CD-ROMs since your system
interpret file name on CD-ROM as only upper case or only lower case, etc.
- Files in the distributed tar-file now unfold into a directory refsys as stated in the manual.
- ANSI-C functions OpenFile and CloseFile in file src/sr.c now exit with value 1 upon error.
All functions in this file print an error on a new fresh line.
Version 0.1, 23/02/1999
This was the first development release. It was circulated to a small group of
people only.