Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Electrical and Computer Engineering


Parsa, Vijay


Accurate speech quality measures are highly attractive and beneficial in the design, fine-tuning, and benchmarking of speech processing algorithms, devices, and communication systems. Switching from narrowband telecommunication to wideband telephony is a change within the telecommunication industry which provides users with better speech quality experience but introduces a number of challenges in speech processing. Noise is the most common distortion on audio signals and as a result there have been a lot of studies on developing high performance noise reduction algorithms. Assistive hearing devices are designed to decrease communication difficulties for people with loss of hearing. As the algorithms within these devices become more advanced, it becomes increasingly crucial to develop accurate and robust quality metrics to assess their performance. Objective speech quality measurements are more attractive compared to subjective assessments as they are cost-effective and subjective variability is eliminated. Although there has been extensive research on objective speech quality evaluation for narrowband speech, those methods are unsuitable for wideband telephony. In the case of hearing-impaired applications, objective quality assessment is challenging as it has to be capable of distinguishing between desired modifications which make signals audible and undesired artifacts. In this thesis a model is proposed that allows extracting two sets of features from the distorted signal only. This approach which is called reference-free (nonintrusive) assessment is attractive as it does not need access to the reference signal. Although this benefit makes nonintrusive assessments suitable for real-time applications, more features need to be extracted and smartly combined to provide comparable accuracy as intrusive metrics. Two feature vectors are proposed to extract information from distorted signals and their performance is examined in three studies. In the first study, both feature vectors are trained on various portions of a noise reduction database for normal hearing applications. In the second study, the same investigation is performed on two sets of databases acquired through several hearing aids. Third study examined the generalizability of the proposed metrics on benchmarking four wireless remote microphones in a variety of environmental conditions. Machine learning techniques are deployed for training the models in the three studies. The studies show that one of the feature sets is robust when trained on different portions of the data from different databases and it also provides good quality prediction accuracy for both normal hearing and hearing-impaired applications.