Speech Quality Assessment using Mel Frequency Spectrograms of Speech Signals

摘要

Non-intrusive speech quality assessment (NI-SQA) has gained importance, due to recent advancements in multimedia, signal processing, machine learning, speech communication, and automatic speech recognition. The performance of NI-SQA techniques highly dependent on the extracted features to predict speech quality. In this article, a new machine learning-based method is proposed for predicting speech quality, without using reference signals is proposed. Traditional techniques used in literature cannot be implemented in practical application scenarios due to less correlation accuracy between subjective and objective scores. In this work, we used Mel-frequency cepstral coefficients (MFCCs) for predicting speech quality that is degraded in different noise conditions. We have computed the proposed work results on two independent databases. Experimental results show significant improvement in the performance when compared with current approaches for assessment of speech quality.