Co-channel speech detection via spectral analysis of frequency modulated sub-bands

Overlapped-speech is known to degrade performance in automatic speech systems. In this study, a sub-band speech analysis technique is proposed to detect overlapped-speech segments in single-channel multi-speaker scenarios (i.e., co-channel speech). Sub-band signals are obtained by decomposing the input speech using a gammatone filterbank. Filterbank outputs are then used to modulate the frequency argument of a sinusoidal carrier. We show that the spectra of these frequency-modulated signals, namely Gammatone Sub-band Frequency Modulation (GSFM) features, are more disperse in overlapped-speech segments compared to single-speaker regions. We quantify the dispersion rate to obtain a measure for the amount of overlapped speech in a given speech segment. Overlap detection experiments are conducted using the speech separation challenge corpus and GSFM features are compared to commonly used overlap detection features. Detection errors are reduced by a relative 50% across different signal-to-interference values ranging from 0 to 9dB.