logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-04162018-154315


Tipo di tesi
Tesi di laurea magistrale
Autore
BENELLI, GIONATA
URN
etd-04162018-154315
Titolo
Artificial Intelligence applied to keyword spotting for low-power and memory constrained embedded device
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
EMBEDDED COMPUTING SYSTEMS
Relatori
relatore Prof. Fanucci, Luca
relatore Palla, Alessandro
Parole chiave
  • artificial intelligence
  • constrained devices
  • deep learning
  • keyword spotting
  • low-power
  • neural network
  • riconoscimento vocale
Data inizio appello
07/05/2018
Consultabilità
Non consultabile
Data di rilascio
07/05/2088
Riassunto
Given a domotic installation based on the Internet of things paradigm, we would like to provide a Voice Wake up to enable hands-free interaction for a user with impaired mobility for increasing quality of life.
We focused on developing a keyword spotting system for constrained devices to enable a low-power voice interaction.
In order to meet the requirements we propose an architecture based on Deep Neural Network. We explore different options based on Convolutional Neural Network, and we propose a new architecture which gives an increase of over 5\% accuracy over standard Convolution architecture, with just a tenth of the total memory footprint if we compare it with previous research on Convolutional Neural Networks for embedded device. We also explore data-augmentations and pseudo-labeling as a way of increasing the robustness of the network without increasing size. Finally we test our implementation over a Raspberry Pi 3,obtaining a reduction of power consumption of over 50\% with respect to previous works, but with the same accuracy.
File