World’s top scientists in speaker recognition together again – this year at Johns Hopkins

The world’s leading researchers in speaker recognition just can’t imagine spending summer without each other. After the success of Brno Speaker Recognition Workshop BOSARIS 2010 and BOSARIS 2012 (well, the 2012 took place a bit later than in summer), the group gathered again – this year hosted by the Center for Language and Speech Processing at the Johns Hopkins University. The workshop takes place from June 24 to July 26 as a continuation of the tradition of JHU summer workshop series.

The people present at JHU (in alphabetical order) are

Jahangir Alam CRIM, Canada Ph.D. student
Hagai Aronowitz IBM Haifa Research Lab, Israel researcher
Niko Brummer Agnitio, Spain senior researcher
Lukas Burget BUT, Czech Republic senior researcher
Sandro Cumani BUT, Czech Republic post-doc researcher
Najim Dehak MIT, USA senior researcher
Patrick Kenny CRIM, Canada senior researcher
David Martínez University of Zaragoza, Spain Ph.D. student
Ignacio Lopez Moreno Google Inc., USA software engineer
Oldrich Plchot BUT, Czech Republic PhD student
Albert du Preez Swart Agnitio, Spain researcher
Themos Stafylakis CRIM, Canada post-doc researcher
Carlos Vaquero Agnitio, Spain post-doc researcher
Karel Vesely BUT, Czech Republic PhD student
Jesús Villalba University of Zaragoza, Spain post-doc researcher

with others involved distantly.

The topics the team tackles include:

  • Automatic clustering of speakers in large data collections
  • Adaptation of speaker recognition system to unlabeled mismatched data sets
  • Accuracy of speaker recognition for short durations, where the phonetic content is the predominant source of intersession variability
  • Speaker recognition robust to distortions (e.g. additive noise)
  • Investigation into how to make artificial neural networks useful in speaker recognition

The team closely cooperates with concurrently running SCALE (Summer Camp for Applied Language Exploration) workshop organized by JHU Human Language Technology Center of Excellence (HLTCoE). SCALE targets “Limited training data”, “Noisy labels” and “Large-scale speaker clustering with side-information” problems - overlapping topics allow the two teams to benefit from intense cooperation.