Analysis of Non-Matching Reference Approach to Predict Speech Intelligibility

This paper takes a critical look at the non-matching reference based speech quality assessment (NORESQA) framework. We posit that the NORESQA framework learns speech intelligibility as a function of the phonetic difference between the input and the reference speech, especially for highly intelligible inputs. To reduce the reliance on phonetic differences, two additional heads are added to the original NORESQA setting. The gender head learns the gender difference between the input and the reference speech, while the speaker head learns the speaker difference between the input and the reference. In terms of ESTOI correlation, modified NORESQA slightly outperforms baseline NORESQA on 3-second recordings, under different training and testing conditions.