Apple: stronger language discrimination narrows the multilingual speech gap

On October 2, 2026, Apple Machine Learning published a paper on the gap between multilingual and monolingual self-supervised speech models.

The paper's central finding is that multilingual self-supervised speech models still fall short of monolingual models when the total pretraining data budget is matched. Multilingual models can benefit from sharing information across languages, but that sharing is not enough to close the gap under a matched-budget comparison.

According to the paper, strengthening the model's ability to discriminate languages during pretraining reduces this multilingual gap. On some measures, it closes the gap altogether, with the multilingual model matching monolingual performance. The measures named in the work include continuous phonetic measures as well as higher-level linguistic measures. The paper also states that cross-language information sharing is retained while the gap narrows.

These conclusions come from a set of English-French bilingual HuBERT controlled experiments. The available material does not report the size of the experiments, the amount of data used, or results for any additional languages, and it does not describe how the approach performs on other languages or other model families. Nothing is said about when the gap appears during training beyond the matched-budget setting, and no figures or baselines are given beyond the comparison itself.