Automatic speaker recognition is critical for many applications, ranging from secure access to intelligence gathering, to archiving and understanding conversation. Current speaker recognition systems model specific speaker characteristics, but a vast range of habitual and stylistic differences has just begun to be explored. These include patterns of intonation, energy and duration, as well as habitual word and phrase usage. Exploiting information in these heterogeneous modes of variation presents challenges in feature selection, modeling, and information combination. Feature discovery and selection efforts will consider the large variety of stylistic features that may be available. The feature space transformation and modeling phase of the work will explore the feature space using dimensionality reduction and clustering. The resulting features will be modeled to focus on specific classes of features. Further system combination research will study how individual systems for specific feature types can best be combined to optimize performance recognition. The new features and modeling approaches will be evaluated in the annual Speaker Recognition Evaluation.
The proposed work will lead to identification of new extractable features characterizing individual speaker behavior. It explores more sophisticated models to better capture complex behavior and relationships. The project has impact for intelligence, law enforcement, security and other application by enhancing recognition performance. Because the new features are based on performance behavior rather than simply vocal tract physiology, the new features can also be used for tasks such as emotion recognition or conversation detection. The systems will be freely available and engage under-represented graduate students.
