The Sensing tasks, such as evaluating voice clarity, are difficult for computers. (SQA). Traditional approaches to SQA have favored objective techniques that require the existence of a separate reference pair. In practice, however, where reliable information about the world is often hard to come by, these approaches unavoidably fall short. Recent months have seen a surge in interest in non-invasive techniques that use neural networks to forecast evaluations or scores; however, these approaches suffer from inconsistency and require labelled data for training. In this piece, we advocate for a fresh approach to evaluating the quality of a speaker's voice. Taking cues from humans' innate ability to compare and evaluate the quality of speech signals even when they have non-matching contents, we introduce a novel framework that predicts a subjective relative quality score for a given speech signal with respect to any given reference without using any subjective data. This score's reliability depends on how well the incoming speech sample matches the standard. We use a convolutive and additive noise-exposed speech dataset that has been assessed using crowd-sourced QoE labels, Pearson association with MOS labels, and mean-squared error of the estimated MOS. The finest Pearson association (0.87) and the lowest mean squared error of any of the techniques we suggest are obtained by a completely connected deep neural network trained on Mel-frequency features. (0:15)

Non-intrusive speech assessment models, deep learning, multi-objective learning, speech enhancement.