WavLM-TDNN cuts speaker verification errors by 24% in classroom noise

In a new arXiv paper, Saba Tabatabaee and colleagues at the University of Maryland tackle a critical problem for AI in education: verifying who is speaking in noisy, real-world classrooms. The team adapted the WavLM-TDNN architecture for speaker verification (SV) using a dataset of English classroom speech with sparse speaker labels. Compared to the widely used ECAPA-TDNN baseline, WavLM-TDNN delivered an average relative Equal Error Rate (EER) reduction of 23.99%, and even beat an ECAPA-TDNN model fine-tuned on classroom data by 6.32%.

The researchers also compared training strategies for low-label environments. They tested pure self-supervised learning (SSL) versus a two-stage approach that first pre-trains with SSL, then fine-tunes on limited annotated data. Across five-fold cross-validation, the two-stage strategy consistently won, achieving a relative EER reduction of 13.39% over SSL alone. These gains held for both children and adult speakers, making the model especially relevant for educational tools like automatic attendance, personalized tutoring, and classroom analytics. The paper is available on arXiv (2608.03623).