ICSI SRE 2004 System Description
-------------------------------------

The submitted system is an implementation of that described in
[1]. Please refer to the paper for greater detail. The system uses
background and target keyword models generated by Hidden Markov Models
(HMMs) for 19 select keywords (where a "keyword" is a word or common
word-pair) drawn primarily from the discourse marker, backchannel and
filled pause categories.  The system employs speaker-independent
keyword-specific HMMs which are then adapted to the target training
data to create target models, and computes test scores using the usual
likelihood ratio of target to background.  Only intervals of speech
corresponding to the 19 keywords are scored.

This system is intended to serve as a contrast to a submission by SRI
which combines scores from this system with those of a more
conventional frame-based GMM system.

Feature Extraction 
==================== 
The HMM feature vectors consist of 19 mel cepstra, the zeroth
cepstrum, and their first differences, for a total of 40 features per
vector. Cepstral Mean Subtraction was performed over the union of speech-rich
segments for each conversation side (segmentation provided by SRI).

Background Model 
==================== 
Keyword UBMs were obtained by training on data from the SRE 2003
Extended Data set (Switchboard II, phases 2 and 3), using splits 6
through 10. We anticipate mismatch with the evaluation data, e.g. due
to the absence of cellular data and the limited demographic
distribution in this set, and intend to examine the effects of using
background models from a better matched training set after the
evaluation.

The keyword HMMs were simple left-to-right state sequences with
self-loops and no skips. Each state model consisted of a mixture of
four gaussians and the number of states for each keyword was defined
to be the smaller of the number of phones in the standard
pronunciation of the word, multiplied by 3, and the median duration in
frames, divided by four. All modeling and scoring was performed using
the HMM Toolkit, HTK.

Training
==================== 
Speaker-specific keyword models were obtained by MAP adaptation of the
background models by adapting only the means of the gaussians. In the
event that there is no training data for a particular keyword, the UBM
is simply copied as the speaker-specific model. This results in
removing the influence of the keyword, as the contribution to the
overall score is zero, due to the cancellation of target and
background. Keyword locations within the audio file were determined by
word-level alignment information made available from SRI's automatic
speech recognition (ASR) system. (See their submission for more detail
on the speech recognizer.)

Testing
==================== 
Each keyword appearing in the test segment is scored by taking the
difference between the log probabilities obtained from scoring the
speaker-specific and UBM models against the test tokens. The final
score is obtained by adding these keyword scores and normalizing by
the total number of frames. Again keyword locations were obtained from
SRI's ASR output.

*Note: For trials involving non-English test segments or having no
 English adaptation data, a score was not computed since no ASR output
 was provided by SRI for these files.  Instead, a dummy value of '0.0'
 was assigned to the trial.

Score Normalization
====================
Standard normalizations such as ZNORM and TNORM have not been included
in this system owing to time constraints, though it is known they may
be able to assist in mitigating the effects of acoustic mismatch. We
intend to examine this following the evaluation.

Computation 
==================== 
All computation was performed on a fleet of Intel 2.8GHz Xeon
Processors with 2GB of memory

The processing times are as follows:

Feature Extraction
----------------------
total elapsed: 302068.00s
total cputime: 33296.20s

1-side/1-side Scoring
----------------------
Training time
total elapsed: 6123.93s
total cputime: 204.64s

Testing time
total elapsed: 130014.00s
total cputime: 25489.00s
  
8-sides/1-side Scoring
----------------------
Training time
total elapsed: 8436.58s
total cputime: 440.54s

Testing time
total elapsed: 53045.90s
total cputime: 16865.00s

*Note:These times do not include the processing required to generate
the ASR output used to locate the keywords. This information can be
found in the SRI system description.

Acknowledgments
====================
We would like to thank SRI for their assistance in this evaluation, esp. 
for providing the ASR output used in this submission.

References
==================== 
[1] K. Boakye and B. Peskin, "Text-Constrained Speaker Recognition on a 
Text-Independent Task," Odyssey 2004: The Speaker and Language Recognition 
Workshop, Toledo, Spain, May/June 2004.