MLA-Project-DYSARTHRIA

Contributors: Bastien DUSSARD, Marc FAVIER, Fabien PICHON, Louis SIMON

Master 2 ISI, Sorbonne Université

Subject: Learning to detect dysarthria from raw speech

Rerefence paper: J. Millet et N. Zeghidour, « Learning to detect dysarthria from raw speech », arXiv:1811.11101 [cs], jan. 2019 Available on: http://arxiv.org/abs/1811.11101

Dataset is available here: http://www.cs.toronto.edu/~complingweb/data/TORGO/torgo.html

---

Python functions and notebooks

The model is build and train using two python files:

• blocks.py which contains elementary blocks of the architecture
• model1.py which gathers model building, training, and a main function

Examples of command:

python model1.py -frontEnd melfilt -normalization pcen -lr 0.001 -batch_size 2 epochs 10 -decay True

python model1.py -frontEnd LLD -lr 0.003 -batch_size 8 epochs 10

python model1.py -frontEnd TDfilt -lr 0.002 -batch_size 4 epochs 10 -decay True

These functions utilize a set of four files which allow to extract and pre-process the data:

• time_mfb.py for Time Domain filterbanks
• LLD_extract.py for LLDs
• melfilt_preprocess.py for mel-filterbanks
• create_dataset.py for raw speech extraction

Finally, one notebook notebook_MLA_dysarthria.ipynb allows to evaluate models.:exclamation: This notebook is design for Google Colab with the GPU on.

Libaries

tensorlflow, librosa and opensmile

Note on Time-domain filterbanks model and GPU

❗ THE USE OF THE GROUPED CONVOLUTION MAKE THE USE OF A GPU MANDATORY. THE BACKPROPAGATION CAN NOT BE APPLIED WITH A CPU, THEN WE CAN ONLY DO INFERENCE WITH A GPU ❗