This toolkit makes available some MATLAB code that may be useful for researchers in the field of automatic speaker detection, particularly those that participate in the yearly NIST Speaker Recognition Evaluations.
These tools may be applicable to many other fields of pattern recognition and machine learning, but prospective users are advised that the current documentation is interspersed with jargon that is specific to speaker detection.
All of the code (MATLAB .m-files) is available for download as FoCal.zip (21 kB).
The functions of the toolkit are introduced in the below-referenced web pages. (The code contains further help in the form of comments.)
- Tools for evaluation of the goodness of speaker detection log-likelihood-ratios.
- Tools for calibration (i.e. converting scores to log-likelihood-ratios).
- Tools for fusion of multiple detection scores.
Fr a deeper understanding of these tools for evaluation and calibration, we recommend that you read:
Niko Brummer and Johan du Preez, “Application-Independent Evaluation of Speaker Detection”, to be published, Computer Speech and Language, 2005. [PDF]
This
version was released in July 2005. Please visit this page again in future to
check for updates or corrections.
Bug reports
and other feedback are welcome. Email me at niko dot brummer at gmail dot
com.
-
Niko
Brummer, July 2005.