Newborns’ cry signals contain valuable information related to the state of the infant. Extracting this information requires a cry detection algorithm able to operate in environments with challenging acoustic conditions, since multiple noise sources, such as interferent cries, medical equipments, and persons may be present. Cry detection is an important facility in both residential and public environments, which can answer to different needs of both private and professional users. In the current dissertation the issue of cry detection in professional and acoustic noisy environments such as Neonatal Intensive care units (NICUs) will be investigate. The research, presented in this thesis, describes the developed approaches for the infant cry detection suitable for NICUs as well as an effective training methodology that does not require labeled data collected in the specific domains of use. In the described approaches the acoustic noise reduction is performed processing multiple audio channels using digital signal processing techniques as well as neural strategies. These approaches use Deep Neural Networks, whose training is conducted on a synthetic dataset created by means of a suitable Acoustic Scene Simulation procedure. The Acoustic Scene Simulation allows the creation of a synthetic dataset that, differently from a real-life dataset, can be acquired without access a NICU. The obtained detection results confirm the goodness of the developed approaches overcoming the performance achieved by the algorithms of the state of art taken as reference and proving that a synthetic dataset can be a useful replacement with respect to a real-life dataset, at least in the early design process. The proposed training methodology permits to lower the interaction with a sensitive environment such as a NICU, to the bare minimum and can be exploited to include changes to the environment as needed, without requiring additional acquisition sessions.
I segnali associati al pianto dei neonati contengono preziose informazioni relative allo stato del bambino. L’estrazione di queste informazioni richiede un algoritmo di rilevazione del pianto in grado di operare in ambienti con condizioni acustiche difficili caratterizzati dalla presenza di fonti di rumore come pianti interferenti, apparecchiature mediche e persone. Il rilevamento del pianto infantile è una funzione importante sia negli ambienti residenziali che in quelli pubblici, in grado di rispondere alle differenti esigenze dei professionisti e degli utenti privati. Nella presente dissertazione viene presentata una indagine riguardo alla problematica questione della rilevazione del pianto infantile in ambienti professionali ed acusticamente rumorosi come le unità di terapia intensiva neonatale (UTIN). La ricerca descritta in questa tesi è volta allo sviluppo di approcci per la rilevazione del pianto adatti alle UTIN, nonché alla definizione di una efficace metodologia di allenamento degli algoritmi che non necessiti di dati raccolti negli specifici domini di utilizzo. Negli approcci descritti, la riduzione del rumore acustico viene eseguita su canali audio multipli con tecniche di elaborazione del segnale digitale e strategie neurali. Questi approcci utilizzano delle reti neurali profonde addestrate su un set di dati sintetico creato mediante un’adeguata procedura di simulazione di scene acustiche, senza la necessità di accedere ad una UTIN. I risultati ottenuti confermano la bontà degli approcci sviluppati superando le prestazioni ottenute dagli algoritmi dello stato dell’arte presi come riferimento, dimostrando che un set di dati sintetico può essere un utile rimpiazzo rispetto ad un set di dati della vita reale. La metodologia proposta per l’allenamento delle reti neurali consente di ridurre al minimo l’interazione con ambienti sensibili come le UTIN e permette di elaborare modifiche dei domini di utilizzo senza richiedere sessioni di acquisizione aggiuntive.
Signal Processing algorithms and Learning Systems for Infant Cry Detection / Ferretti, Daniele. - (2019 Mar 13).
Signal Processing algorithms and Learning Systems for Infant Cry Detection
FERRETTI, DANIELE
2019-03-13
Abstract
Newborns’ cry signals contain valuable information related to the state of the infant. Extracting this information requires a cry detection algorithm able to operate in environments with challenging acoustic conditions, since multiple noise sources, such as interferent cries, medical equipments, and persons may be present. Cry detection is an important facility in both residential and public environments, which can answer to different needs of both private and professional users. In the current dissertation the issue of cry detection in professional and acoustic noisy environments such as Neonatal Intensive care units (NICUs) will be investigate. The research, presented in this thesis, describes the developed approaches for the infant cry detection suitable for NICUs as well as an effective training methodology that does not require labeled data collected in the specific domains of use. In the described approaches the acoustic noise reduction is performed processing multiple audio channels using digital signal processing techniques as well as neural strategies. These approaches use Deep Neural Networks, whose training is conducted on a synthetic dataset created by means of a suitable Acoustic Scene Simulation procedure. The Acoustic Scene Simulation allows the creation of a synthetic dataset that, differently from a real-life dataset, can be acquired without access a NICU. The obtained detection results confirm the goodness of the developed approaches overcoming the performance achieved by the algorithms of the state of art taken as reference and proving that a synthetic dataset can be a useful replacement with respect to a real-life dataset, at least in the early design process. The proposed training methodology permits to lower the interaction with a sensitive environment such as a NICU, to the bare minimum and can be exploited to include changes to the environment as needed, without requiring additional acquisition sessions.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Ferretti.pdf
accesso aperto
Descrizione: Tesi_Ferretti.pdf
Tipologia:
Tesi di dottorato
Licenza d'uso:
Creative commons
Dimensione
114.42 MB
Formato
Adobe PDF
|
114.42 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.