Statistical model-based objective measures of speech quality

Good objective measures of speech quality are highly desirable and beneficial in the design and maintenance of diverse speech processing devices and speech communication systems. Normally, objective measures of speech quality are divided into two groups: intrusive and non-intrusive measures. Intrusive measures assess speech quality based on the input and output signals of systems under test, while non-intrusive measures assess speech quality based only on the output signal of systems under test. In this thesis, we developed novel methods for these two types of measurements by applying statistical models.; For the non-intrusive measurement, a continuous-mapping based method was proposed based on an adaptive neuro-fuzzy inference system, in which we applied neural network and fuzzy logic techniques to speech quality evaluation. In addition, we pro posed an innovative pattern classification-based measure, in which hidden Markov models and Bayesian inference were employed to estimate speech quality in line with subjective listening opinion tests. These two measures have been tested and compared with the state-of-the-art non-intrusive measurement standard; the ITU-T P.563, which was published in 2004. The experimental results showed that the correlation coefficients of these two methods for a set of speech databases attained 0.8812 and 0.8393; respectively. These results compared favorably with the ITU P.563, which provided an averaged correlation result of 0.8422.; For the intrusive measurement, a novel loudness pattern distortion measure was proposed, in which we applied the Moore and Glasberg's loudness model to speech quality evaluation. In order to effectively map the extracted loudness information into speech quality ratings, we developed a cognitive model by using Bayeisan learning and Markov chain Monte Carlo methods, in which the model complexity issue was handled in a natural and consistent way. The effectiveness of the proposed method was demonstrated through comparisons with the state-of-the-art intrusive measurement standard, the ITU-T P.862.1 (PESQ+), which was published in 2003. The experimental results showed that the averaged correlation coefficient of the proposed measure attained 0.9524, which compared favorably to the ITU P.862.1 giving a correlation result of 0.9518.; In Keywords. Speech quality; Objective measures; Intrusive measurement; Non-intrusive measurement; Adaptive neuro-fuzzy inference system; Gaussian mixture density; Hidden Markov model; Bayesian inference; Loudness patterns; Equivalent rectangular bandwidth; Bayesian learning; Markov chain Monte Carlo methods.