RANNIS Infrastructure grant 2018 – Awarded.

The complexity of neural network models increases every year and it takes a lot to keep up with computational hardware fast enough to train them efficiently.  Earlier this year, we applied to the RANNIS Infrastructure Fund for funding to expand our current HPC cluster.  We are happy that our proposal “Deep Learning Infrastructure for Speech and Language Technology” was selected to be funded. Only three grants were granted to Reykjavik University, and ours was one of them. This money will allow us to buy a fully equipped SuperMicro 4028GR-TR2 server with NVIDIA 1080Ti GPUs. We hope to sign the grant contract within several weeks and then order the machine. Next comes the process of assembling and integrating it into our current cluster. I can say that several group members can’t wait to have more power available to them.

 

02I72646
List of awarded grants – RANNIS

Article accepted to IEEE/ACM Transactions on Audio, Speech and Language Processing.

On 28/9/2017, we got an article accepted to the special issue of IEEE/ACM Transactions on Audio, Speech and Language Processing. The article is a result of our collaboration with Daryush D. Mehta and Jarrad Van Stan from the Eye and Ear, Massachusetts General Hospital. The title of the article is “Modal and nonmodal voice quality classification using acoustic and electroglottographic features.” It analyses a set of glottal source, vocal tract and harmonic model features in the task of voice quality classification. In a nutshell, we are trying to predict whether the speaker will speak in a breathy, normal, strained or rough voice. The figure below summarizes the achieved results.

res_COVAREP

The initial manuscript was sent for peer-review in December 2016. The special issue is expected to come out by the end of this year and the links to the paper will be available in a few weeks, hopefully :). For this paper, we used recordings of normal people mimicking the asked voice quality so our next target is to extend the results to a much bigger database, which contains recordings of patients with various voice disorders. So stayed tuned, if things go well, we might make our way to Seoul in 2018 for the ICASSP conference with new results.

Happy pictures of all authors below (Michal, Daryush, Jarrad, Jón).

 

Link : IEEE/ACM Transactions on Audio, Speech and Language Processing

Preprint version

ICASSP 2017 New Orleans – we will be there.

Two members of our lab will attend ICASSP 2017 conference in New Orleans. We will present a paper on the classification of voice modalities using a neck-surface accelerometer. The paper is the result of our collaboration with guys from the Massachusetts General Hospital, US and Universidad Tecnica Federico Santa Maria, Chile. Thanks for your help!

Hopefully, the trip won’t be all work and no fun.