This extensible technology isolates data (text, image, and voice) representing a target class from heterogeneous data representing multiple data categories of the same type. The method may be applied to identify speech from one speaker in audio containing several other speakers, and extends to language and gender identification and image and text applications. By auto-selecting data representing a particular class from multi-class data, nonessential artifacts may be removed from models trained on multi-class data, thereby enhancing detection and identification capabilities.
Potential applications include: speech applications (speaker/language recognition, gender identification), improved classification accuracy in pattern recognition, and biometric data discrimination.
- In general, this technology models of one class of data from a selection containing multiple classes of data within a particular category
- Application specific, the system generates a model of one speaker from data containing multiple speakers in which the total speech time of the one speaker exceeds that of any other speaker in the data
- The system can further separate male speech from female speech and different languages or dialects all with a background of other voices
- US patent 7,454,337 available for license
- Potential for collaboration with NSA scientists and engineers