Success of AMI partners in NIST 2006 Speaker recognition evaluation

AMI partners BUT (Czech republic) and TNO (Netherlands) recorded important success in NIST speaker recognition evaluations organized in spring 2006. Their system scored among the best in tough competition of 37 academic and industrial laboratories from all over the world (NIST rules prohibit to disclose the exact position of their system).

Pavel Matejka (BUT), David van Leeuwen (TNO), Niko Brummer (Spescom Data Voice), Petr Schwarz and Lukas Burget (both BUT) on Puerto-Rico beach.

The system was developed within "STBU" consortium - BUT and TNO in cooperation with South-African partners: company Spescom Data Voice and Stellenbosch University. It included a combination of 3 acoustic classification techniques: Gaussian mixture models (GMM) classifying directly speech features, Support vector machines (SVM) processing super-vectors of GMM means and a third sub-system based on SVM-classification of MLLR adaptation matrices from LVCSR system.

Great care was given to transmission channel compensation and score normalization: the successful system includes techniques such as feature mapping, eigen-channel adaptation and nuisance attribute projection (NAP). The scores are normalized by classical t-norm technique and fusing of systems was performed using logistic linear regression.

Speaker identification and recognition has primary use in security and defense (for example for detecting a suspect in hundreds of telephone recordings) but is also useful for mining information in huge audio archives of meetings, or TV and radio broadcasts.

Work on the system was primarily supported by European AMI project (IST 6th framework programme) with contribution of Czech Ministry of Defence and Dutch Ministry of Defence. Powerful computing resources in Brno were supported by CESNET - non-profit organization unifying Czech Universities, responsible for high-speed University backbone and for research in network- and media-related technologies.

The developers of the system are: Pavel Matejka, Lukas Burget, Petr Schwarz and several students from BUT, David van Leeuwen from TNO, Niko Brummer from SDV and Albert Strasheim from Stellenbosch University.

Online papers