Wide Incremental learning with Discrimination nEtworks

01.09.2017 bis 31.08.2022

This five-year project aims to deepen our understanding of how we produce and understand words in everyday speech.
Words in day-to-day conversational speech may differ substantially from how they appear in writing: German "würden" is often pronounced as "wün," Dutch "natuurlijk" ('naturally') can reduce to "tk", and Mandarin 要不然 (jao pu zan, 'otherwise') to “ui." Current theories assume that the sound waves that reach our ears are reduced to sequences of abstract sound units, much like the sequences of letters that make up written words. However, how to align highly reduced forms such as "wün", "tk" and "ui" with their full unreduced variants, the supposed gatekeepers to meaning, is an unsolved computational problem.
The WIDE project makes the radical proposal to eliminate letter-like sound units altogether, and instead to zoom in on the rich details of the speech signal itself. Given tens of thousands of smart features representing the richness of the speech signal, it is anticipated that artificial neural networks can learn, by trial and error, to identify which meanings are conveyed. Previous research funded by the Alexander von Humboldt foundation allowed to provide a first proof of concept. In the WIDE project, this approach will be developed further and extended from German to other languages, including Mandarin Chinese (a tone language) and Estonian (a complex language with 28 to 40 different forms for a given noun). The WIDE project also targets a computational model without sound units for the articulation of words in speech production.
The project's name, "WIDE", highlights a second aspect in which this project makes a radical departure from current trends in linguistics and natural language processing. Instead of making use of deep learning networks, the project focuses on the potential of ‘wide' two-layer networks with tens of thousands of input and output units.