Influence of simultaneous spoken sentences on the properties of spectral peaks

In this study, an approach to analyse the properties of spectral peaks of simultaneously talking speakers in monophonic audio signal has been described. We have proposed a technique based on spectral peaks tracking and attributes calculated from peaks histogram. Spectral peaks have been estimated using linear prediction-based spectral envelope for each frame of source signal. The features have been computed from the histogram at different frequency bands. The statistical properties of the obtained features have been used to find out the relationship with the number of speech sources. Proposed approach has been tested using a dedicated database featuring sentences with the same and mixed gender, where the number of speakers varies from two to twelve. Different configuration parameters like frame size, bin width of the histogram and linear prediction order have been used in the conducted experiments. The results show that obtained trends of statistical descriptors are directly connected with the number of voice sources. The proposed descriptors and performed regression analysis can be a basis to estimate the number of speakers in single audio stream.