EURASIP Journal on Information Security
CNNMC: A Convolutional Neural Network with Monte Carlo Dropout for Speaker Recognition.
Finally accepted at EURASIP Journal on Information Security (Q2, IF 2.1)
Pushing CNNs Beyond the Usual Limits!
We’re excited to share our latest work presenting a new CNN architecture with a custom Monte Carlo Dropout => CNNMC.
We’ve been experimenting with a new CNN using a new customized Monte Carlo Dropout. The CNNMC use this Dropout not only in the training phase, but also during inference.
Why does this matter?
This new CNNMC architecture:
– captures more information.
– reduces variance.
– makes decisions more robust and stable with unseen and noisy data.
The application: Speaker verification and Speaker identification.
The results: higher accuracy for identification task and lower EER for verification task, at inference time compared to other SOTA models: CNN, GMM, LDA, X-vector, I-vector, KNN.
We measured performances with 4 different datasets: BioVID, TIMIT, LibriSpeech, VoxCeleb.
Our Dataset: https://www.dmi.unict.it/spata/BiovidChallenge/
I’m sincerely grateful to all co-authors for the valuable collaboration and for the opportunity to work together on this research.
