Enhancing objective evaluation of speech quality algorithm: current efforts, limitations and future directions

Abstract

AbstractWe discuss a performance improvement over the current state‐of‐the‐art objective speech quality assessment algorithm by means of a parameter study. The state‐of‐the‐art algorithm computes a certain measure of disturbances surface, and applies a cognitive model to map the computed value to the scale used mean opinion score (MOS). We note that many parameters, including the used metric of Lp space, do not appear to be robust enough in terms of the performance of the algorithm. First, we search for optimum Lp‐norms over the frequency‐‐time domain disturbance surface. The optimum Lp‐norms yield the most desirable correspondence between symmetric/asymmetric disturbance terms and subjective scores. New features are added to the cognitive processing algorithm. Improvements are substantial, but still only incrementally improve the performance. Limitations and bottlenecks of the current standardised approaches, as well as emerging new ideas, are reviewed. Finally, we call for more innovative, rather than renovative, research efforts in objective speech quality measurement for fundamental enhancement of the algorithm. Copyright © 2009 John Wiley & Sons, Ltd.