Speech enhancement models are typically trained on randomly sampled mini-batches. The authors of a new arXiv preprint argue that such random sampling can yield batches with limited diversity in the intrinsic acoustic characteristics of clean speech, which may hold back performance given earlier evidence linking those characteristics to enhancement outcomes.

To address this, they propose a feature-stratified mini-batch sampling strategy. Training examples are divided into quantiles according to a chosen clean-speech characteristic, and each mini-batch is constructed to contain equal representation from every quantile. This guarantees within-batch diversity without requiring any modification to the enhancement model or its training objective.

Experiments on two large-scale English and Spanish datasets, plus a smaller dataset containing neurotypical and pathological speech, showed consistent improvements over standard random mini-batch sampling across multiple objective enhancement metrics. Ablation studies indicated the gains were robust across training runs, though their magnitude depended on which acoustic characteristic was used for stratification.

The method is simple and complementary to existing training pipelines, offering a low-cost way to improve speech enhancement performance by changing only how batches are composed.