IDENTIVOX © : A PC-WINDOWS TM TOOL FOR TEXT-INDEPENDENT SPEAKER RECOGNITION IN FORENSIC ENVIRONMENTS
Joaquín González-Rodríguez, Javier Ortega-García, José Juan Lucena-Molina · 2001
A state-of-the-art Gaussian-Mixture-Modelling (GMM) text-inde- pendent speaker recognition system has been developed. The system, perfectly suited to the Bayesian approach for Forensic Speaker Recognition, works in a user-friendly platform (Win95/98/NT/2000/me) in a fully user-configurable envi- ronment. Speaker modelling, identification and verification are the main functions of the system. The system optionally performs channel normalisation and/or likeli- hood normalisation, the latter through a Universal Background Model. The work of the system user is organised in sessions (files), where the user can configure, save, open, and print reports for each individual session. User-selected speech audio files, parameterised into LPCC or MFCC, are used to train the speaker models, and test speech segments are evaluated against the selected set of candidates (identification) or against a claimed model (verification). Some excellent results are provided with different speech databases. A free English-demo version for evaluation of the system is available to forensic and research institutes writing to [email protected].