Open Access

Gauri Deshpande

• This thesis delves into the realm of speech representation and deep learning techniques to extract breathing patterns from speech signals. Breathing patterns—the signals generated during respiration—are intricately connected to speech production. The respiratory organs contribute to the production of speech signals as well, and hence both breathing patterns and speech have an impact on each other. In this thesis, time-domain speech representation, coupled with phase-domain decomposed speech components, is investigated as a carrier of respiratory information. This feature set and a novel long-short-term-memory (LSTM)-based deep architecture are introduced to extract the breathing patterns from the speech signals. The speech-breathing data from 100 healthy college going students, while they read a phonetically balanced text is collected to build this model. The thesis also explores the impact of breathing pattern categories on the performance of the deep model as well as the variabilityThis thesis delves into the realm of speech representation and deep learning techniques to extract breathing patterns from speech signals. Breathing patterns—the signals generated during respiration—are intricately connected to speech production. The respiratory organs contribute to the production of speech signals as well, and hence both breathing patterns and speech have an impact on each other. In this thesis, time-domain speech representation, coupled with phase-domain decomposed speech components, is investigated as a carrier of respiratory information. This feature set and a novel long-short-term-memory (LSTM)-based deep architecture are introduced to extract the breathing patterns from the speech signals. The speech-breathing data from 100 healthy college going students, while they read a phonetically balanced text is collected to build this model. The thesis also explores the impact of breathing pattern categories on the performance of the deep model as well as the variability of model performance observed across the 100 speakers. Furthermore, the pre-trained model is utilised to extract breathing patterns from speech data labelled with respiratory disorders and human-confidence levels. The resulting speech-derived breathing patterns serve as a pioneering feature set for detecting respiratory disorders and gauging human-confidence levels. Expanding on the potential applications of this representation technique, the thesis suggests exploring its use in the domains of physiology and psychology. Specifically, it highlights the opportunity for early diagnosis of a spectrum of respiratory disorders and the assessment of psychological states and traits. This research opens doors to leveraging speech-derived breathing patterns for advancing diagnostic capabilities in respiratory health and understanding psychological aspects.…