Statistical and model based approach to unvoiced speech detection

The detection of unvoiced speech in the presence of additive background noise is complicated by the fact that unvoiced speech is very similar to white noise. The mechanism of production of unvoiced speech is known to be due to turbulent airflow in the constrictions of the vocal tract. Three approaches for detecting unvoiced speech from additive background noise have been developed. Two of which are very effective in the presence of additive white noise, are model based and autocorrelation based respectively. The probability of correct detection, on average being 74%. A statistical approach is however developed that works both for additive white and pink noise. Further research on this statistical measure is being attempted to use it in a simple threshold based detector of unvoiced speech.