Voicing-aware parametric speech quality models over VoIP networks

摘要

This paper describes novel parametric speech quality models which subsume the effect of packet loss distribution and voicing feature of missing signal waves. Speech quality estimate models for voiced and unvoiced loss location patterns are developed following multiple statistical regression analysis of measurements gathered from a built speech quality assessment framework. The overall speech quality is estimated by combining voiced and unvoiced speech quality estimate scores using an expression calibrated using a large number of speech samples. The input parameters namely, mean loss durations and ratios for voiced and unvoiced packets, of speech quality estimate models are extracted at run-time using a new voicing-aware packet loss Markov model. This chain, calibrated at run-time, finely models bursty packet loss behavior over voiced and unvoiced missing speech waves. Performance evaluation study shows that our voicing-aware speech quality estimate models clearly outperform voicing-agnostic speech quality models in terms of accuracy over a wide range of conditions.