Vector quantization techniques for output-based objective speech quality

Output-based speech quality (OBQ) refers to objective speech quality assessment using only received speech without utilizing the input speech record. This paper proposes three new OBQ measures and evaluates their performance. Parameters derived from perceptual linear prediction (PLP) coefficients are used to provide speaker independence required by the objective measures. PLP, PLP cepstrum, and PLP delta-cepstrum parameters are computed for output speech records from an undistorted source speech database and vector quantized. The resulting codebook provides a reference for computing objective distance measures for distorted speech. The proposed objective measures are the transition probability distance, the median minimum distance, and the chi-squared distance. The OBQ parameters are tested on four different speech datasets, and correlation is computed between subjective scores and objective distances under a variety of conditions. The results indicate that the proposed algorithms are robust against speaker, text, and distortion variation