Working memory (WM) is often poorer for a second language (L2). In low noise conditions, people listening to a language other than their first language (L1) may have similar auditory perception skills for that L2 as native listeners, but do worse in high noise conditions, and this has been attributed to the poorer WM for L2. Given that WM is critical for academic success in children and young adults, these speech in noise effects have implications for academic performance where the language of instruction is L2 for a student. We used a well-established Speech-in-Noise task as a verbal WM (vWM) test, and developed a model correlating vWM and measures of English proficiency and/or usage to scholastic outcomes in a multi-faceted assessment medical education program. Significant differences in Speech-Noise Ratio (SNR50
) values were observed between medical undergraduates who had learned English before or after five years of age, with the latter group doing worse in the ability to extract whole connected speech in the presence of background multi-talker babble (Student-t tests, p < 0.001). Significant negative correlations were observed between the SNR50
and seven of the nine variables of English usage, learning styles, stress, and musical abilities in a questionnaire administered to the students previously. The remaining two variables, Perceived Stress Scale (PSS) and the Age of Acquisition of English (AoAoE) were significantly positively correlated with the SNR50
, showing that those with a poorer capacity to discriminate simple English sentences from noise had learnt English later in life and had higher levels of stress – all characteristics of the international students. Local students exhibited significantly lower SNR50
scores and were significantly younger when they first learnt English. No significant correlation was detected between the SNR50
and the students’ Visual/Verbal Learning Style (r = −0.023). Standard multiple regression was carried out to assess the relationship between language proficiency and verbal working memory (SNR50
) using 5 variables of L2 proficiency, with the results showing that the variance in SNR50
was significantly predicted by this model (r2 = 0.335). Hierarchical multiple regression was then used to test the ability of three independent variable measures (SNR50
, age of acquisition of English and English proficiency) to predict academic performance as the dependent variable in a factor analysis model which predicted significant performance differences in an assessment requiring communications skills (p = 0.008), but not on a companion assessment requiring knowledge of procedural skills, or other assessments requiring factual knowledge. Thus, impaired vWM for an L2 appears to affect specific communications-based assessments in university medical students.