Extracting Video-Based Breath Signal For Detection of Out-of-breath Speech

Abstract

A cost-effective video signal based breath signal extraction method is described in this work. It does not require any sophisticated instrument; instead uses devices like mobile phones, headphones and computers that are readily available to an individual. For the same, a new database is created having read-speech utterances and video signals under the neutral and the post-exercise (or known as out-of-breath) conditions. The breath signals for most of the speakers exhibit a higher strength for both inhalation and exhalation phases of the breathing cycle under out-of-breath conditions. Additionally, the average duration of the breath cycle decreases for the same condition. The exhalation phase mainly influences the above time reduction. The ability of the breath features for distinguishing the neutral and the out-of-breath class is verified by the support vector machine and the logistic regression classifiers. The performance of both the classifiers in terms of unweighted average recall and Fl-score improved to $\approx$ 70% after combining the above breath features with the MFCC baseline features.