An Investigation of Automatic Phonetic-Unit Selection for Forensic Voice Comparison

A hybrid Hidden Markov Model (HMM) - Gaussian Mixture Model (GMM) system was proposed to automatically select tokens of /iau/, /ai/, /ei/, /m/ and /n/ in a database of recordings of Standard-Chinese speech collected under studio-clean, mobile-landline degraded and mismatched recording conditions. The FVC systems constructed were all MFCC GMM-UBM systems, but based on different portions of the recordings. Fusion of an FVC system based on portions of the recordings within manual /iau/ markers with a baseline system based on all speech-active segments of the recordings resulted in a relatively large improvement in validity in all three conditions. Index Terms: forensic voice comparison, automatic phonetic unit selection, validity and reliability