Miloš Cerňak (Idiap Research Institute, Martigny, Switzerland) “Phonological posteriors and their applications”



The phonological class probabilities are the probabilities of individual phonological classes (known also as phonological or distinctive features) inferred from a short segment of speech signal. A vector of all phonological class probabilities is referred to as phonological posterior. We use the term phonological posteriors as an analogy to well-known phone posteriors. In this talk, I will introduce converging evidence about phonological posteriors from linguistics and neuroscience, which is a basis of our motivation in studying and using them. Phonological posteriors have two basic properties: (i) the binary nature, where most of their values are very close to either 0 or 1, and (ii) although they are extracted as segmental features, their structures (patterns) convey supra-segmental (linguistic and prosodic) information as well. I will demonstrate their usability on a wide range of applications, namely 1) comparison of phonological systems, 2) the linguistic parsing, 3) the prosodic even detection, 4) progressive apraxia of speech markers, 5) phonological speech synthesis and 6) parametric speech coding. Thus, phonological posteriors might be used as unique features, or as complementary features to phone posteriors.


Milos holds a PhD in Computer science from Slovak University of Technology in Bratislava. Since 2011 he is a Scientific Collaborator at Idiap Research Institute, Martigny, Switzerland, involved in various Swiss R&D academic and industrial projects with a core focus on speech recognition and synthesis, low bit rate speech coding, and prosody parametrization. From 2008, he was a member of IBM Research in Prague, working on embedded ViaVoice speech recognition and IBM Reading Companion. After graduating in 2005, he was a post-doc researcher at Institute EURECOM in France, and a principal researcher at Slovak Academy of Sciences. During his doctoral studies in 2001, he was also a visiting scientist at Iowa State University’s Virtual Reality Application Centre, Ames, USA. He is interested in signal processing, phonology, and neural basis of speech production and perception.

Center for Language and Speech Processing