A measure of aperiodicity content in a speech signal

Abstract

Most of the current aperiodicity detectors measure aperiodicity indirectly, in which the absence of periodic component in a nonsilent region is termed as aperiodicity. Such indirect measurements are inadequate and can be misleading especially in cases like voiced fricatives or breathy vowels. This motivated us to develop a direct measure of aperiodicity which is independent of the periodicity measure. The speech signal is passed through a 60-channel gamma tone auditory filterbank. Average Magnitude Difference Function (AMDF) is computed on the envelope of each channel. The randomness in the distribution of the AMDF dips is the basis for the measure of the aperiodicity whereas the measure of the periodicity is based on the occurrence of the dips at multiple locations. The system was evaluated on the MOCHA database, which has simultaneous recording of the EGG data, and on the TIMIT database. Preliminary analysis shows that the aperiodicity and voicing accuracy on a per frame basis are 95% and 90.3%, respectively. In voiced fricatives and voiced stops, high aperiodicity and high periodicity were detected in 21.6% of the frames. Note that not all these sounds necessarily had both the sources prominent.