Perceptually-based objective quality using phoneme-level segmentation

Research in objective speech-quality incorporates perceptual aspects of human hearing such as nonlinear frequency sensitivity, loudness, and loudness level to improve the accuracy of quality estimates. Not even these new algorithms, however, are sufficiently reliable to replace human listener panels. While such hearing models include physical effects of the human ear, they do not account for higher-order processing in the brain such as phoneme-, word-, and sentence-level analysis. The paper studies the effects of phoneme-level distortion on subjective quality. Subjectively rated speech records were segmented into related phoneme groups, and processed using perceptually-based distortion measures. The database includes both waveform and low bit-rate nonwaveform coders. Estimated correlation coefficients indicate the relative importance of each phoneme group to perceived quality. Regression coefficients are used to weight phoneme distortions, which are combined into an overall objective quality score. The paper compares test results for perceptually-based objective quality with and without phoneme-level segmentation