NNArt2020: Neural network models for articulatory gestures University of British Columbia Vancouver, Canada, July 9, 2020 |
Conference website | https://staff.fnwi.uva.nl/t.o.lentz/nnart/ |
Submission link | https://easychair.org/conferences/?conf=nnart2020 |
Abstract registration deadline | March 15, 2020 |
Submission deadline | March 15, 2020 |
Satellite workshop at LabPhon 17
Purpose
Articulation has been formalised as dynamic articulatory gestures, i.e., a target-driven pattern of articulator movements (e.g., Browman & Goldstein, 1986). Such a pattern unfolds in time and space and could therefore also be seen as a spatial sequence of analytically relevant articulatory landmarks such as timepoint of peak velocity and target achievement. Seeing such sequences as sequences of vectors (of spatial coordinates) make them potentially learnable with algorithms for sequence modelling.
Current developments of machine learning offer greatly improved power for sequence learning and prediction. Recurrent Neural Networks (RNNs) or their extension Long Short-Term Memory (LSTM, Hochreiter & Schmidhuber, 1997) allows efficient training over short and even long time intervals (Gers, Schraudolph & Schmidhuber, 2002). Such networks have been used for acoustic modelling, but their application in articulation research has been mainly been limited to ultrasound data, and applied less to the classification of two-dimensional articulator movement curves as obtained from EMA or ROI analyses of MRI data.
However, promising approaches to acoustics-to-EMA mapping tentatively suggest that articulatory movement allow meaningful modelling using deep neural networks (e.g., Liu et al., 2005, Chartier et al., 2018)
This workshop aims at bringing together researchers interested in such modelling and to foster research collaborations (also sharing of datasets) . Possible topics for papers are:
- Whether it is possible to capture invariants, language-independent predictable patterns that apply to all articulation
- If transfer learning is possible, i.e. if a network trained on articulatory features in one speaker and, ultimately, language can be mapped onto the pattern of another speaker (or language)
- If annotation of gestures can be aided by generating most likely gesture structures, analogous to the derivation of articulation from acoustics (e.g., Mitra, Vikramjit, et al. 2010)
- If diagnostic classification is possible on networks that model articulation, analogous to e.g., the detection of counterparts to compositionality in a model of arithmetic grammar by Hupkes & Zuidema (2017)
We envisage an informal workshop, with talks followed by discussion. Length of talks will depend on the number of submissions. We will end with a synthesis session, in which the issue of data formats, data sharing and algorithm sharing are addressed.
Submission Guidelines
Submissions should follow the following conference guidelines (the same as for the main conference):
- Abstracts should be written in English and not exceed one page of text. References, examples and/or figures can optionally be included on a second page.
- Submitted abstracts must be in .pdf format, with Times New Roman font, size 12, 1 inch margins and single spacing
Abstracts do not have to be anonymous.
Contact
See for more information https://staff.science.uva.nl/t.o.lentz/nnart/