A Missing-Data Approach to Noise-Robust LPC Extraction for Voiced Speech Using Auxiliary Sensors
Cenk Demiroğlu, Thomas P. Barnwell III · 2006
Noise robust LPC extraction from the voiced speech signal is addressed with a missing-data approach. Harmonics in the voiced speech spectrum are detected using a general electromagnetic motion sensor (GEMS) that is immune to acoustic background noise. Non-harmonic frequencies are treated as missing-data and severely suppressed while no processing is done on the harmonic frequencies since they are assumed to have high SNR. Objective measure tests using the log likelihood ratio (LLR) show significant improvement over the noisy case for severely noisy environments.