Usefulness of text-conditioning and a new database for text-dependent speaker recognition research

Text Dependent (TD) Speaker Recognition systems assume that the password to be uttered by the speaker is known to the system. As the password is known, the system can apply a password-specific model capturing the speaker dynamics well. This enables TD systems to perform better than textindependent systems. We present a variation of the TD systems, called text-conditioning, in which the password is uniquely chosen by each user. This delivers a higher level of discrimination since the linguistic and phonetic differences of the passwords themselves are exploited in separating the speakers. As the database for such a study was not publicly available, we built an extensive database for speaker recognition having such text-conditioning property. The database is tested with various speaker recognition trials. The results indicate that for the design of a practical TD speakerrecognition system, “text-conditioning” does offer a significant edge.