Nov. 27, 2022, 11:47 a.m.

In a few days, the world's leading neural networks conference NeurIPS 2022 will be held in New Orleans.

A key co-author of the first of the papers Variable-rate hierarchical CPC leads to acoustic unit discovery in speech (pdf) is our alumnus Santiago Cuervo. The publication was prepared during his studies in our Data Science Programme and has five authors: Santiago Cuervo, Adrian Łańcucki [NVidia], Ricard Marxer [Université de Toulon, Aix Marseille Univ, CNRS, LIS, France], Paweł Rychlikowski (II UWr) and Jan Chorowski [Pathway, France].

A method for discovering phonemes (i.e., components of speech) is described, in a situation where we only have unscripted recordings of speech in a given language. Particularly important in the method developed was the ability to find phonemes of different lengths, which better corresponds to the characteristics of speech, and which was not present in earlier methods. This research can be placed in the context of the more ambitious goal of discovering not just phonemes, but words, their meaning, grammar and the language as a whole system, when we only hear utterances in the language, devoid of any context and not written down as text in any way.

The second paper to appear at this conference is FlowHMM: Flow-based continuous hidden Markov models (pdf) and is authored by Paweł Lorek (IM UWr, Tooploox), Rafał Nowak (II UWr, Tooploox), Tomasz Trzciński (PW, UJ, Tooploox, IDEAS) and Maciej Zięba (PWr, Tooploox).

Furthermore, at the recently concluded leading conference in theoretical computer science FOCS 2022, a paper Cut query algorithms with star contraction was published by Pawel Gawrychowski and his co-authors S. Apers [Paris], Y. Efron [Columbia], T. Lee [Sydney], S. Mukhopadhyay [Sheffield], and D. Nanongkai [Copenhagen].


New Orleans skyline
(fot. thepipe26 CC BY 2.0)