Breath detection during speech has broad applications ranging from emotion recognition to detection of diseases. Most of the breath detection equipment are contact based. In the proposed method, we use a voice activity detector (VAD) to find the non-speech region and searches the breath only in this region since breath is a non-speech activity. This reduces the execution time. A support vector machine (SVM) classifier is used with radial basis function (RBF) kernel trained on the cepstrogram feature to detect the breaths in the non-speech regions. The classifier output is post-processed to join breathing segments which are closely spaced and remove small duration breaths. Speech breathing rate is calculated as the ratio of the number of breaths to the time between the first and last breath. The algorithm is tested on a student evaluation database. The algorithm yields an F1 Score of 94% and root mean square error (RMSE) of 7.08 breaths/min for the speech-breathing rate. The output has been validated using thermal videos. The breaths have been classified as full and partial detection based on the Intersection over Union (IOU). The algorithm is also tested on some news channel reports which gave a minimum F1 Score of 73%.
Cite as: K., M.I.Y.A., Routray, A. (2019) Automatic Detection of Breath Using Voice Activity Detection and SVM Classifier with Application on News Reports. Proc. Interspeech 2019, 609-613, doi: 10.21437/Interspeech.2019-2434
@inproceedings{k19_interspeech,
author={Mohamed Ismail Yasar Arafath K. and Aurobinda Routray},
title={{Automatic Detection of Breath Using Voice Activity Detection and SVM Classifier with Application on News Reports}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={609--613},
doi={10.21437/Interspeech.2019-2434}
}