Segmentation of Speech and Humming in Vocal Input

Abstract

Non-verbal vocal interaction (NVVI) is an interaction method in which sounds other than speech produced by a human are used, such as humming. NVVI complements traditional speech recognition systems with continuous control. In order to combine the two approaches (e.g. "volume up, mmm") it is necessary to perform a speech/NVVI segmentation of the input sound signal. This paper presents two novel methods of speech and humming segmentation. The first method is based on classification of MFCC and RMS parameters using a neural network (MFCC method), while the other method computes volume changes in the signal (IAC method). The two methods are compared using a corpus collected from 13 speakers. The results indicate that the MFCC method outperforms IAC in terms of accuracy, precision, and recall.