Kubichek Developing effective automatic, or objective techniques for assessing speech quality to replace human listener scores (i.e., subjective quality) has been the object of much research. Current algorithms base quality estimates on input-to-output distortion measures. A related, yet almost unexplored problem, is estimation of transmission quality using only received speech without access to the transmitted speech record (Output-Based Quality or OBQ). A second problem is objectively measuring intelligibility using only received speech (Output-Based Intelligibility or OBI). This research addresses both of these difficult problems, and is based on recently developed technologies that utilize models of hearing perception to provide speaker-independent speech recognition. OBQ and OBI estimates are being determined from carefully designed distance measures between Perceptual Linear Prediction (PLP) coefficients of output speech and cluster centroids derived from training data. A PLP analysis system is being implemented. Methods for estimating OBQ and OBI estimates are being developed. Finally, the algorithms are being tested on a variety of speech databases. Statistical analyses will characterize their performance in terms of their correlation with subjective results and robustness to speaker and distortion variation.
