Pre-trained neural models for detecting respiratory insufficiency in speech

• Grant number: 20/16543-7
• Support type: Scholarships in Brazil - Post-Doctorate
• Effective date (Start): February 01, 2021
• Effective date (End): June 30, 2022
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal researcher: Marcelo Finger
• Grantee: Marcelo Matheus Gauy
• Home Institution: Instituto de Matemática e Estatística (IME). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated research grant: 20/06443-5 - Spira: system for early-detection of respiratory insufficiency by voice audio analysis, AP.R

Abstract

This project was conceived in the context of the construction of a system for the automated learning of the detection of respiratory insufficiency through the analysis of audio signal from the patients. This proposal aims to build pre-trained models and process audio processing in Portuguese, which will later be used for specific training, such as classes specification of the audio being delivered by a person who needs or who does not need hospitalization for respiratory failure. Pre-trained models have two phases. In a first phase a very large amount of data is used for non-training supervised; this phase is called pre-training. On the second phase, the model is training specifically for an application, with a much smaller amount of data; this phase is called training specific development. Pre-trained models have obtained good results in word processing; we intend to develop metrics that allow for checking the quality of pre-trained audio models.