Grants Speaker state recognition

Many researchers lead their efforts to analyze signals from diverse sources to discover the different states that a person can manifest. The concept of “speaker states” can be applied to emotional status, psychological considerations, alcohol intoxication or sleepiness, among others. One of the main motivations is the wide variety of practical applications of great interest, among which could be named medical care, security, entertainment, etc. In this project new models, algorithms and computational tools based on pattern recognition are developed, which are applicable to speech analysis in order to infer the different states that people can manifest. Different techniques of signal processing and perceptual or bio-inspired models will be investigated to achieve the best sets of characteristics for the different states present in the speakers. With respect to the classification techniques, specific hierarchical and ensemble models will be developed in order to take advantages of particularities of each task and to improve the recognition rates achieved by state-of-the-art systems. In addition, both representations and classifiers will be evaluated in signals with different noises to verify their robustness. In addition, new challenges as the classification using multiple corpus and/or ​​never-seen languages will be addressed.

Enrique Marcelo Albornoz