Developing a robust technique for denoising and extracting speech and breath sounds in voice recordings from asthmatic patients

Abstract

Auscultation and processing cough, voice and breath sounds play an important role in diagnosis of several pulmonary ailments. There have been a number of studies using machine learning algorithms on such sound files to build classification and prediction algorithms. Since these studies used specialized microphones in controlled environments, it is difficult to test and deploy these algorithms in real-life settings. Recorded speech files consist of breath and wheeze sounds and it is challenging to extract from this single sound file. Hence, several audio processing and editing software are used to demarcate these sounds. The proposed technique uses a combination of a denoiser and an extraction technique to overcome these drawbacks. The developed pipeline ensures that the audio files are free of any environmental and background noises, and the audio can be recorded through any kind of microphone and environmental settings. The extraction technique further is the result of combinations of filters to output the speech and breath sounds as individual sound files, ready for processing and eliminating the need of audio editing and processing software.