|
|
Audio Classification with Deep Learning on Limited Data Sets
Harár, Pavol ; Platoš,, Jan (oponent) ; Šimák, Boris (oponent) ; Mekyska, Jiří (vedoucí práce)
Standard procedures of dysphonia diagnosis by a clinical speech therapist have their downsides, mainly because the process is very subjective. Recently, an automatic objective analysis of a speaker's condition gained in popularity. Researchers successfully based their methods on various machine learning algorithms and handcrafted features. These methods, unfortunately, are not directly scalable to other voice disorders and the process of feature engineering is laborious and thus financially and talent expensive. Based on the previous successes, a deep learning approach might help to ease the problems with scalability and generalization, but an obstacle is a limited amount of training data. This is a common denominator in almost all systems for automated medical data analysis. The main aim of this work is to research new approaches to deep-learning-based predictive modeling using limited audio data sets, focusing especially on voice pathology assessment. This work is the first to experiment with deep learning in this field and on so far the largest combined database of dysphonic voices, which was created in this work. It provides a thorough examination of publicly available data sources and identifies their limitations. It describes the design of novel time-frequency representations based on Gabor transform and it presents a new class of loss functions, that yield target representations beneficial for learning. In numerical experiments, it demonstrates improvements in the performance of convolutional neural networks trained on limited audio data sets using the augmented target loss function and the newly proposed time-frequency representations, namely Gabor and Mel scattering.
|
|
|
Audio Classification with Deep Learning on Limited Data Sets
Harár, Pavol ; Platoš,, Jan (oponent) ; Šimák, Boris (oponent) ; Mekyska, Jiří (vedoucí práce)
Standard procedures of dysphonia diagnosis by a clinical speech therapist have their downsides, mainly because the process is very subjective. Recently, an automatic objective analysis of a speaker's condition gained in popularity. Researchers successfully based their methods on various machine learning algorithms and handcrafted features. These methods, unfortunately, are not directly scalable to other voice disorders and the process of feature engineering is laborious and thus financially and talent expensive. Based on the previous successes, a deep learning approach might help to ease the problems with scalability and generalization, but an obstacle is a limited amount of training data. This is a common denominator in almost all systems for automated medical data analysis. The main aim of this work is to research new approaches to deep-learning-based predictive modeling using limited audio data sets, focusing especially on voice pathology assessment. This work is the first to experiment with deep learning in this field and on so far the largest combined database of dysphonic voices, which was created in this work. It provides a thorough examination of publicly available data sources and identifies their limitations. It describes the design of novel time-frequency representations based on Gabor transform and it presents a new class of loss functions, that yield target representations beneficial for learning. In numerical experiments, it demonstrates improvements in the performance of convolutional neural networks trained on limited audio data sets using the augmented target loss function and the newly proposed time-frequency representations, namely Gabor and Mel scattering.
|
host ::
RSS.